Robot Automatic Obstacle Avoidance Target Recognition System and Method

Abstract

This invention discloses an automatic obstacle avoidance target recognition system and method for robots, relating to the field of autonomous navigation technology. The method includes: acquiring and fusing obstacle boundary maps, passable area masks, and dynamic target heatmaps to output an environmental feature map; inputting the environmental feature map into a pre-trained deep reinforcement learning model to generate an obstacle avoidance instruction set based on the robot's real-time state and global path planning results; collecting actual robot path data and comparing it point-by-point with the expected path; if the persistent deviation feature exceeds a preset persistent deviation feature threshold, updating the environmental feature map, obstacle avoidance instruction set, and persistent deviation feature by inputting them into the deep reinforcement learning model. This invention improves the obstacle avoidance capability of autonomous robots in dynamic and complex environments through multimodal perception fusion and a closed-loop dynamic optimization mechanism.

Description

Technical Field

[0001] This invention relates to the field of autonomous navigation technology, and in particular to a robot automatic obstacle avoidance target recognition system and method. Background Technology

[0002] Currently, autonomous mobile robots are widely used in industrial logistics, intelligent inspection, and other fields. Their obstacle avoidance capabilities directly determine the safety and efficiency of task execution. In dynamic and complex environments, robots need to perceive multi-dimensional environmental information in real time and make accurate decisions. Traditional methods rely on single sensor data or static environmental models, making it difficult to effectively integrate multi-modal information such as the spatial distribution of obstacles, traversable areas, and the movement trends of dynamic targets.

[0003] Especially in unstructured scenarios, existing technologies cannot achieve spatiotemporal continuity modeling of multi-source perception data and joint optimization of dynamic obstacle avoidance strategies, resulting in problems such as delayed avoidance response, path planning oscillations, and sudden target tracking failures. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a target recognition method for automatic obstacle avoidance in robots to solve the problems of insufficient multimodal perception fusion and poor real-time obstacle avoidance decision-making in dynamic and complex environments.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] In a first aspect, the present invention provides a robot automatic obstacle avoidance target recognition method, which includes acquiring an obstacle boundary map, a passable area mask and a dynamic target heat map and fusing them to output an environmental feature map;

[0008] The environmental feature map is input into a pre-trained deep reinforcement learning model, and an obstacle avoidance instruction set is generated based on the robot's real-time state and global path planning results.

[0009] Based on the obstacle avoidance instruction set, predict the remaining width of the avoidance space, identify ground no-turn signs, predict the dynamic target trajectory, dynamically adjust the steering angle and braking level, and output the corrected obstacle avoidance instruction set.

[0010] The modified obstacle avoidance instruction set is converted into control signals to drive the robot to perform obstacle avoidance actions;

[0011] Collect the robot's actual path data and compare it point by point with the expected path. If the continuous deviation feature exceeds the preset continuous deviation feature threshold, input the environmental feature map, obstacle avoidance instruction set and continuous deviation feature into the deep reinforcement learning model for updating.

[0012] In a preferred embodiment of the robot automatic obstacle avoidance target recognition method of the present invention, the specific steps for outputting the environmental feature map are as follows:

[0013] Calculate the dynamic confidence weight of each sensor based on the real-time noise parameters of the multimodal sensor;

[0014] Based on dynamic confidence weights, the obstacle boundary map, passable area mask, and dynamic target heat map are aligned to a unified topological coordinate system;

[0015] The obstacle boundary map, passable area mask, and dynamic target heat map are fused to output a fused spatiotemporal feature map;

[0016] Verify the consistency of the fused spatiotemporal feature maps based on the preset map topology, and output the environmental feature map.

[0017] In a preferred embodiment of the robot automatic obstacle avoidance target recognition method of the present invention, the specific steps for outputting the fused spatiotemporal feature map are as follows:

[0018] The obstacle boundary map, passable area mask, and dynamic target heat map are mapped to the joint manifold space to generate spatiotemporally unified multimodal projection data;

[0019] Spatiotemporal tensor decomposition is performed on multimodal projection data to separate the principal components of the motion trend of dynamic targets and the principal components of the structure of static environment;

[0020] By selecting the dominant feature subspace, it is reconstructed into a fused spatiotemporal feature map.

[0021] In a preferred embodiment of the robot automatic obstacle avoidance target recognition method of the present invention, the specific steps for generating the obstacle avoidance instruction set are as follows:

[0022] By using a neural symbolic joint encoding network, environmental feature maps, robot real-time states, and global path planning results are aligned across modalities in time and space to generate dynamic knowledge-enhanced feature vectors.

[0023] Embed a differentiable symbolic logic layer in a pre-trained deep reinforcement learning model to generate original obstacle avoidance instructions based on feature vectors enhanced by dynamic knowledge.

[0024] A dynamic knowledge graph is constructed based on the original obstacle avoidance instructions. Conflicts between instructions and physical rules are detected through graph reasoning, triggering a quantum-inspired instruction recombination algorithm to eliminate the conflicts.

[0025] An incremental online distillation mechanism is adopted to synchronously update the recombined instructions with real-time sensor feedback to generate an obstacle avoidance instruction set.

[0026] In a preferred embodiment of the robot automatic obstacle avoidance target recognition method of the present invention, the specific steps for outputting the corrected obstacle avoidance instruction set are as follows:

[0027] Based on the robot's real-time speed and turning angle, dynamically predict the remaining width of the avoidance space and generate a safety marker for the avoidance space.

[0028] The system uses a visual camera to identify ground no-turn signs. If a ground no-turn sign is detected and the confidence level exceeds a preset confidence threshold, a ground constraint sign is generated.

[0029] Kalman filtering is used to predict the trajectory of dynamic targets, calculate the overlap rate with the robot path, and generate dynamic conflict markers.

[0030] Adjust the steering angle and braking level based on the combined status of obstacle avoidance space safety signs, ground constraint signs, and dynamic conflict signs;

[0031] Adjust parameters based on actual obstacle avoidance path deviation, and output the corrected obstacle avoidance command set.

[0032] In a preferred embodiment of the robot automatic obstacle avoidance target recognition method of the present invention, the specific steps for driving the robot to perform obstacle avoidance actions are as follows:

[0033] Based on the steering angle and braking level in the revised obstacle avoidance instruction set, execution priorities are assigned through a dynamic behavior tree to generate motor steering and braking pulse control signals.

[0034] Based on dynamic conflict markers and clearance space safety markers, dynamic interference intensity levels are defined, and adaptive pulse width modulation is used to adjust the signal duty cycle.

[0035] The physical feasibility of the control signal is verified through real-time simulation, invalid signals are filtered out and the timing signal is reconstructed;

[0036] The servo motor and brake are driven by the reconstructed timing signals to synchronously execute steering and braking actions.

[0037] The execution status is monitored in real time through an ultra-low latency feedback channel. If signal drift is detected, a self-repair protocol is triggered to regenerate the control signal.

[0038] In a preferred embodiment of the robot automatic obstacle avoidance target recognition method of the present invention, the input deep reinforcement learning model is updated, and the specific steps are as follows.

[0039] The robot's actual path data is collected in real time by fusion positioning of lidar and odometry, and the path deviation sequence is generated by comparing it point by point with the expected path.

[0040] Perform dynamic trend analysis on the path deviation sequence to extract persistent deviation characteristics;

[0041] If the persistent deviation feature exceeds the preset persistent deviation feature threshold, the current environment feature map, obstacle avoidance instruction set and persistent deviation feature are associated to construct incremental training samples;

[0042] The parameters of the deep reinforcement learning model are dynamically updated based on incremental training samples.

[0043] Secondly, the present invention provides a robot automatic obstacle avoidance target recognition system, including an environment fusion module, a strategy generation module, a dynamic adjustment module, a control execution module and a closed-loop optimization module;

[0044] The environment fusion module is used to acquire obstacle boundary maps, passable area masks, and dynamic target heat map, and fuse them to output an environment feature map;

[0045] The strategy generation module is used to input the environmental feature map into a pre-trained deep reinforcement learning model and generate an obstacle avoidance instruction set based on the robot's real-time state and global path planning results.

[0046] The dynamic adjustment module is used to predict the remaining width of the obstacle avoidance space based on the obstacle avoidance instruction set, identify ground no-turn signs, predict the trajectory of the dynamic target, dynamically adjust the steering angle and braking level, and output the corrected obstacle avoidance instruction set.

[0047] The control execution module is used to convert the modified obstacle avoidance instruction set into control signals to drive the robot to perform obstacle avoidance actions;

[0048] The closed-loop optimization module is used to collect the robot's actual path data and compare it point by point with the expected path. If the continuous deviation feature exceeds the preset continuous deviation feature threshold, the environmental feature map, obstacle avoidance instruction set and continuous deviation feature are input into the deep reinforcement learning model for updating.

[0049] Thirdly, the present invention provides a computer device including a memory and a processor, wherein the memory stores a computer program, wherein when the computer program is executed by the processor, it implements any step of the robot automatic obstacle avoidance target recognition method as described in the first aspect of the present invention.

[0050] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the robot automatic obstacle avoidance target recognition method as described in the first aspect of the present invention.

[0051] The beneficial effects of this invention are as follows: By using multimodal perception fusion and closed-loop dynamic optimization mechanisms, the obstacle avoidance capability of autonomous robots in dynamic and complex environments is improved; by using cross-modal fusion of multimodal data, a spatiotemporally correlated global environmental feature representation is constructed; by using differentiable modeling to analyze the continuous motion trend of dynamic targets, the trajectory prediction accuracy of sudden obstacles is enhanced; by combining the dynamic reward mechanism of deep reinforcement learning with physical constraint fusion strategies, rapid response and motion smoothness are synergistically optimized; and by using incremental learning driven by closed-loop feedback, the robot autonomously adapts to unknown environmental disturbances, forming a full-process obstacle avoidance architecture with autonomous evolution capabilities. Attached Figure Description

[0052] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0053] Figure 1 A flowchart for a robot's automatic obstacle avoidance target recognition method.

[0054] Figure 2 This is a block diagram of a robot's automatic obstacle avoidance and target recognition system.

[0055] Figure 3 A flowchart generated for the environmental feature map.

[0056] Figure 4 A flowchart generated for the obstacle avoidance instruction set. Detailed Implementation

[0057] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0058] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0059] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0060] Reference Figures 1-4This is one embodiment of the present invention, which provides a robot automatic obstacle avoidance target recognition method, including the following steps:

[0061] S1. Obtain obstacle boundary map, passable area mask and dynamic target heat map and fuse them to output environmental feature map.

[0062] Furthermore, based on the real-time noise parameters of the multimodal sensors, the dynamic confidence weight of each sensor is calculated.

[0063] It should be noted that the real-time noise parameters of the LiDAR sensor include point cloud density and occlusion rate, the real-time noise parameters of the visual camera sensor include image blur and feature matching success rate, and the real-time noise parameters of the infrared sensor include the signal-to-noise ratio of the heat source signal. Through dynamic confidence weight calculation, the real-time noise parameters of the LiDAR, visual camera, and infrared sensor are mapped to a unified noise evaluation index. Combined with exponential decay, the contribution weight of each sensor in environmental feature fusion is dynamically adjusted to ensure that the data influence of high-noise sensors is effectively suppressed, while the data proportion of low-noise sensors is increased.

[0064] Specifically, based on the real-time noise parameters of the multimodal sensors, the dynamic confidence weight of each sensor is calculated, as expressed by:

[0065] ;

[0066] In the formula, Indicates the first Sensor-like sensors at the current time Dynamic confidence weights, Indicates the sensor index. Indicates the current time. Indicates the first The dynamic attenuation coefficient of the sensor-like sensor, Indicates the first Sensor-like sensors at the current time Real-time confidence parameters, This indicates that the summation is performed across all sensor types. This represents the dynamic attenuation coefficient across all sensor types. This indicates that all sensor types are being traversed. This indicates iterating through all sensor types at the current time. The real-time confidence parameter.

[0067] It should be noted that the sensor types include LiDAR sensors, vision camera sensors, and infrared sensors. The dynamic attenuation coefficient of the sensor is calibrated by the physical characteristics of each sensor (such as anti-interference ability and measurement accuracy) and historical scene data, such as in a conventional structured environment (corridor / indoor). =0.2 (LiDAR sensor), =0.5 (visual camera sensor), =0.4 (infrared sensor); The dynamic attenuation coefficient for all sensor types is modeled through multi-sensor synergy (e.g., complementarity analysis) and calibrated using normalization constraints, for example, the value range is... .

[0068] Based on dynamic confidence weights, the obstacle boundary map, passable area mask, and dynamic target heat map are aligned to a unified topological coordinate system;

[0069] Specifically, the obstacle boundary map generated by the lidar sensor, the passable area mask generated by the vision camera sensor, and the dynamic target heat map generated by the infrared sensor are spatially registered. The obstacle boundary map is mapped to Euclidean distance field through the affine transformation matrix of the topological coordinate system, the passable area mask is subjected to morphological dilation and erosion operations, and the dynamic target heat map is subjected to trajectory interpolation under the constraints of spatiotemporal continuity. The feature maps of the three are weighted and superimposed using dynamic confidence weights to achieve pixel-level spatial alignment under a unified topological coordinate system.

[0070] The obstacle boundary map, passable area mask, and dynamic target heat map are fused to output a fused spatiotemporal feature map;

[0071] It should be noted that the fused spatiotemporal feature map includes,

[0072] The obstacle boundary map, passable area mask, and dynamic target heat map are mapped to the joint manifold space to generate spatiotemporally unified multimodal projection data;

[0073] Specifically, the obstacle boundary map is mapped to the static geometric subspace of the joint manifold space through orthogonal basis function decomposition. The passable region mask is embedded into the feasible region probability subspace of the joint manifold space using a probability density diffusion algorithm. The dynamic target heat map is projected to the dynamic evolution subspace of the joint manifold space through spatiotemporal tensor decomposition. Based on dynamic confidence weights, the manifold coordinates of the three subspaces are weighted and fused under non-Euclidean metrics. The topological structure of each subspace is uniformly calibrated using Lie group operations in spatiotemporal differential manifolds. The dimensional offset error of cross-modal projection is eliminated through manifold curvature adaptive constraints, generating spatiotemporally aligned multimodal joint projection data.

[0074] Spatiotemporal tensor decomposition is performed on multimodal projection data to separate the principal components of the motion trend of dynamic targets and the principal components of the structure of static environment;

[0075] It should be noted that the principal components of the motion trend of the dynamic target are modeled using the time derivative term in the spatiotemporal integral term and exponentially decaying weights to capture the instantaneous velocity and acceleration characteristics of the dynamic target in the local coordinate system. The partial derivative of the spatiotemporal tensor with respect to time is calculated using historical trajectory data within a time window. This is combined with an exponentially decaying function to strengthen the weights of recent motion states and suppress the influence of outdated trajectory information, generating principal components of the dynamic target's motion trend in the horizontal and vertical directions. These components characterize the future position prediction and avoidance space compression rate of moving obstacles.

[0076] The structural principal components of the static environment are extracted through orthogonal basis function decomposition, mapping the probability distribution of the Euclidean distance field and the passable region mask of the obstacle boundary map to a spatial modal subspace composed of horizontal and vertical basis functions. Singular value decomposition is used to separate the global channel topology from the local obstacle contours, generating the structural principal components of the static environment in the horizontal and vertical directions, characterizing the channel width, curve curvature, and geometric constraints of the feasible region boundaries.

[0077] Specifically, the spatiotemporal tensor decomposition of the multimodal projection data is expressed as follows:

[0078] ;

[0079] In the formula, Represents the multimodal environment feature tensor. Represents horizontal spatial coordinates. Represents vertical spatial coordinates. Indicates multimodal channel, Represents a time variable. Indicates the time window Accumulate points. Indicates the time variable Time position The original spacetime tensor at that location, Indicates the time variable Time position The original spacetime tensor at that location versus time variable The partial derivatives, Indicates the attenuation coefficient. Indicates the current time With time variables Time difference, Represents the integral of the time variable. This represents the tensor product operation. Principal component index representing the structure of a static environment. This represents the total number of principal components in the static environment. This represents the summation of the principal components of the structure over all static environments. Indicates the first Singular values ​​of the principal components of a static environment Represents horizontal spatial coordinates The Orthogonal basis functions of the principal components of a static environment Representing vertical spatial coordinates The Orthogonal basis functions of the principal components of a static environment The weighting function represents the manifold curvature term. Represents trace operation. Represents the curvature tensor. Represents the three-dimensional curvature tensor. Indicates the x-axis direction. Indicates the y-axis direction. express Axial direction.

[0080] It should be noted that the attenuation coefficient is matched with the sensor characteristics based on the dynamic scene response requirements and optimized through experimental calibration. For example, the value range in a normal environment is 0.2~0.4.

[0081] By selecting the dominant feature subspace, it is reconstructed into a fused spatiotemporal feature map.

[0082] It should be noted that the preset singularity threshold is set based on the cumulative variance contribution rate of the multimodal projection data. Principal component analysis is used to calculate the variance proportion of each modality's singularities, and a threshold exceeding the cumulative variance contribution rate is used as the screening criterion. For example, the preset singularity threshold for the LiDAR sensor modality is set to 0.5, for the visual camera sensor modality to 0.3, and for the infrared sensor modality to 0.4.

[0083] Specifically, principal component analysis is used to screen the dominant feature subspaces with singular values ​​greater than a preset singular value threshold. The spatiotemporal differential terms of the principal components of the dynamic target motion trend and the orthogonal basis function terms of the principal components of the static environment structure are retained. Dynamic confidence weights are used to weight and fuse the principal components of the dynamic target motion trend and the principal components of the static environment structure. The fused spatiotemporal feature map is then reconstructed by combining the multimodal projection data of spatiotemporal tensor decomposition.

[0084] Verify the consistency of the fused spatiotemporal feature maps based on the preset map topology, and output the environmental feature map.

[0085] It should be noted that the preset map topology is jointly defined by channel connectivity constraints, minimum obstacle spacing constraints, and feasible path continuity constraints. Among them, the channel connectivity constraints require that the minimum connection width of adjacent feasible regions meets preset conditions. The preset conditions refer to the minimum connection width of adjacent feasible regions, the minimum Euclidean distance between static obstacles, and the maximum curvature derivative of the rate of change of curvature of feasible paths, which are used to verify the consistency of the fused spatiotemporal feature maps.

[0086] Specifically, based on the channel connectivity constraints and obstacle minimum spacing constraints defined by the preset map topology, the geometric distribution of obstacles in the fused spatiotemporal feature map is compared with the obstacle positions in the preset map topology using Euclidean distance fields. The probability distribution of the feasible region generated by the passable area mask is matched with the probability density of feasible paths in the preset map topology. The trajectory prediction generated by the dynamic target heatmap is used to detect trajectory conflicts with the dynamic rules in the preset map topology. The spatial registration error of the obstacle boundary map, passable area mask, and dynamic target heatmap is adjusted by dynamic confidence weights. The iterative nearest point algorithm is used to optimize local coordinate offsets. The dimensional offset of cross-modal projection is eliminated by manifold curvature constraints. After multiple rounds of verification, the environmental feature map is output.

[0087] S2. Input the environmental feature map into the pre-trained deep reinforcement learning model, and generate an obstacle avoidance instruction set based on the robot's real-time state and global path planning results.

[0088] Furthermore, by using a neural symbolic joint encoding network, the environmental feature map, the robot's real-time state, and the global path planning results are aligned across modalities in time and space to generate a dynamic knowledge-enhanced feature vector.

[0089] It should be noted that the neural symbolic joint encoding network consists of a deep feature encoder and a symbolic logic inferencer. The deep feature encoder takes the environmental feature map as input and extracts multi-scale spatial features through a convolutional neural network; the symbolic logic inferencer, based on prior rules in a dynamic knowledge graph, performs symbolic encoding of the robot's real-time state and global path planning results under logical constraints.

[0090] Specifically, through a neural symbolic joint encoding network, the geometric constraints of obstacles, the probability distribution of feasible regions, and dynamic trajectory predictions in the environmental feature map are aligned with the robot's real-time pose, velocity, and acceleration using a spatiotemporal tensor under dynamic confidence weights. The path curvature, steering angle constraints, and target point coordinates of the global path planning results are mapped to the symbolic logic space of the neural symbolic joint encoding network through Lie group manifold operations. Trajectory conflict rules and topological connectivity rules from a dynamic knowledge graph are used to jointly encode cross-modal features under logical constraints, calibrating the kinematic relationship between the robot's real-time state and the environmental feature map, and generating dynamic knowledge-enhanced feature vectors.

[0091] Embed a differentiable symbolic logic layer in a pre-trained deep reinforcement learning model to generate original obstacle avoidance instructions based on feature vectors enhanced by dynamic knowledge.

[0092] It should be noted that by loading historical expert trajectory datasets into the simulation environment for imitation learning, and using a behavior cloning algorithm to map the action-state pairs demonstrated by experts to the output probability distribution of the policy network, a course learning strategy is adopted to gradually increase the number of dynamic obstacles and environmental complexity. Multiple expert policies are fused into unified policy network initialization parameters through policy distillation. Combined with a meta-learning framework, the policy network is pre-trained in multi-task scenarios to generate pre-trained weights with cross-scenario generalization capabilities, which are then used for pre-training the deep reinforcement learning model.

[0093] Specifically, the differentiable symbolic logic layer is deployed in parallel with the policy network of the deep reinforcement learning model through a differentiable interface. Dynamically enhanced feature vectors are input into both the policy network and the differentiable symbolic logic layer. The policy network extracts experience-driven behavioral policies based on the deep reinforcement learning model, while the differentiable symbolic logic layer parses the dynamically enhanced feature vectors based on the obstacle avoidance rule base in the dynamic knowledge graph, generating rule-based obstacle avoidance instructions under symbolic logic constraints. The Q-value output of the policy network and the rule confidence of the differentiable symbolic logic layer are weighted and fused through the differentiable interface. Dynamic confidence weights are used to adjust the contribution ratio of experience policies and logical rules to generate the original obstacle avoidance instructions.

[0094] A dynamic knowledge graph is constructed based on the original obstacle avoidance instructions. Conflicts between instructions and physical rules are detected through graph reasoning, triggering a quantum-inspired instruction recombination algorithm to eliminate the conflicts.

[0095] Specifically, based on the semantic segmentation results of the original obstacle avoidance commands, spatial semantic labels are extracted for obstacle boundaries, passable areas, and dynamic target areas. These semantic labels are mapped to entity nodes and relational edges in a dynamic knowledge graph, constructing a graph topology that includes obstacle locations, path connectivity, and dynamic trajectory constraints. A graph inference engine is used to retrieve physical rules in the dynamic knowledge graph that conflict between path planning commands and the spatial distribution of obstacles and the movement trends of dynamic targets. When a conflict is detected between a command and the minimum safe distance rule or the maximum curvature constraint, a quantum-inspired command recombination algorithm is triggered to encode the conflicting command into a quantum superposition state. A quantum entanglement gate is used to generate a joint command state, and Pauli measurements are used to filter command combinations that satisfy the Lyapunov stability condition, outputting an optimized command sequence that resolves the conflict.

[0096] An incremental online distillation mechanism is adopted to synchronously update the recombined instructions with real-time sensor feedback to generate an obstacle avoidance instruction set.

[0097] Specifically, an incremental online distillation mechanism is employed. The recombined instructions output by the quantum-inspired instruction recombination algorithm, along with real-time feedback data from LiDAR, visual cameras, and infrared sensors, are input into a dynamic knowledge graph. The graph inference engine retrieves conflict patterns between the instructions and the real-time obstacle distance field, feasible region probability distribution, and dynamic target trajectory. Fuzzy membership functions in the differentiable symbolic logic layer are used to calculate the instruction confidence weights. Based on these confidence weights, online knowledge distillation is performed on the recombined instructions and sensor feedback. The Q-value distribution in the policy network is updated using dynamic confidence weights. A Lie group manifold is used to calibrate the spatiotemporal evolution direction of the recombined instructions and the spatial registration error of the sensor feedback. After iterative optimization, an obstacle avoidance instruction set is generated.

[0098] S3. Based on the obstacle avoidance instruction set, predict the remaining width of the avoidance space, identify ground no-turn signs, predict the trajectory of dynamic targets, dynamically adjust the steering angle and braking level, and output the corrected obstacle avoidance instruction set.

[0099] Furthermore, based on the robot's real-time speed and turning angle, the remaining width of the avoidance space is dynamically predicted, and a safety marker for the avoidance space is generated.

[0100] Specifically, based on the robot's real-time speed and turning angle, the distance field of the feasible region mask edge in the environmental feature map is extracted. Combined with the minimum safe distance constraint of obstacles in the dynamic knowledge graph, the lateral compression gradient of the feasible channel under the current pose is calculated in real time by interpolation, and the remaining width of the avoidance space is dynamically predicted. The evolution trend of the remaining width is verified by the curvature derivative constraint in the spatiotemporal differential manifold. When the compression gradient is detected to exceed the Lyapunov stability boundary, the generation of the avoidance space safety marker is triggered.

[0101] The system uses a visual camera to identify ground no-turn signs. If a ground no-turn sign is detected and the confidence level exceeds a preset confidence threshold, a ground constraint sign is generated.

[0102] It should be noted that the preset reliability threshold is based on the joint statistical analysis of the historical recognition accuracy of the visual camera and the distribution of ambient lighting conditions. The balance between the recognition accuracy and false detection rate of the visual camera is optimized through experimental calibration; for example, the preset reliability threshold for ground no-turn signs is set to 0.7.

[0103] Specifically, RGB images captured by a vision camera are input into a convolutional neural network that fuses multi-scale features. A sliding window mechanism is used to extract local texture and shape features of no-turn signs on the ground, matching them with no-turn sign categories in a ground sign template library. When the confidence score of the detection result exceeds a preset confidence threshold, the pixel coordinates of the recognition box are transformed to the robot's local coordinate system. Combined with path connectivity rules in a dynamic knowledge graph, a geofence boundary for the no-turn area is generated. The geofence boundary is then mapped to ground constraint signs in the environmental feature map through spatiotemporal tensor decomposition.

[0104] Kalman filtering is used to predict the trajectory of dynamic targets, calculate the overlap rate with the robot path, and generate dynamic conflict markers.

[0105] Specifically, a dynamic target trajectory prediction method using Kalman filtering is employed. The real-time position and velocity of the dynamic target are used as state variables, and a sequence of trajectory points within a future time window is predicted through a state transition equation. Based on the robot path planning results, path curvature and coordinate point distribution are extracted. The minimum Euclidean distance between the predicted trajectory points and path points is calculated, and the percentage of trajectory points with a distance less than that in the dynamic knowledge graph is used as the overlap rate. When the overlap rate exceeds a preset conflict threshold in the dynamic knowledge graph, a dynamic conflict marker is generated.

[0106] It should be noted that the preset conflict threshold is based on the robot kinematic constraints and the minimum safe distance between dynamic targets defined in the dynamic knowledge graph, combined with the path connectivity rules and the historical statistical characteristics of the dynamic target's motion trend, and the optimal balance point is determined through Monte Carlo simulation testing. For example, in the dynamic obstacle avoidance test scenario, the preset conflict threshold is set to 30%, that is, when the overlap rate between the Kalman filter predicted trajectory and the robot path exceeds 30%, a dynamic conflict flag is triggered.

[0107] Adjust the steering angle and braking level based on the combined status of obstacle avoidance space safety signs, ground constraint signs, and dynamic conflict signs;

[0108] Specifically, the obstacle avoidance safety marker triggers a braking level gradient enhancement mechanism based on Lyapunov stability conditions; the ground constraint marker suppresses the feasible region boundary of the steering angle through a quantum-inspired command reorganization algorithm; and the dynamic conflict marker uses spatiotemporal tensor decomposition to predict the trajectory conflict region and generate steering angle compensation. Lie group manifold operations are used to map the constraints of the three types of markers to the joint control command space. Multi-source command weights are fused through fuzzy membership functions of a differentiable symbolic logic layer to output steering angle and braking level parameters.

[0109] Adjust parameters based on actual obstacle avoidance path deviation, and output the corrected obstacle avoidance command set.

[0110] Specifically, based on the Euclidean distance field deviation between the actual obstacle avoidance path and the planned path, a parameter adjustment mechanism is triggered through the path connectivity rule base in the dynamic knowledge graph. The stability condition of the path deviation is calculated using a Lyapunov function, generating dynamic confidence weights and steering angle compensation amounts. An incremental online distillation mechanism is used to backpropagate the gradient information of the path deviation to a quantum-inspired instruction reorganization algorithm, adjusting the quantum superposition state weight distribution of the steering instructions. The fuzzy membership function of the differentiable symbolic logic layer verifies whether the corrected steering angle and braking level meet the minimum safe distance constraint in the dynamic knowledge graph, outputting the corrected obstacle avoidance instruction set.

[0111] S4. Convert the modified obstacle avoidance instruction set into control signals to drive the robot to perform obstacle avoidance actions.

[0112] Furthermore, based on the steering angle and braking level in the revised obstacle avoidance instruction set, execution priorities are assigned through a dynamic behavior tree to generate motor steering and braking pulse control signals.

[0113] Specifically, the dynamic behavior tree evaluates the joint state priority of avoidance space safety markers, ground constraint markers, and dynamic conflict markers through condition nodes, and calculates the execution order of steering angle adjustment commands and braking level enhancement commands using dynamic confidence weights. The path connectivity rule base in the dynamic knowledge graph defines the action mapping relationship of the behavior tree execution nodes. The highest priority steering angle command is converted into the duty cycle of the motor steering pulse through Lie group manifold operations, and the braking level command is mapped into the amplitude and timing of the braking pulse through Lyapunov stability constraints. The node weights of the dynamic behavior tree are updated through an incremental online distillation mechanism to generate motor steering and braking pulse control signals.

[0114] Based on dynamic conflict markers and clearance space safety markers, dynamic interference intensity levels are defined, and adaptive pulse width modulation is used to adjust the signal duty cycle.

[0115] Specifically, based on dynamic conflict identifiers and avoidance space safety identifiers, dynamic interference intensity levels are defined through the path connectivity rule base in the dynamic knowledge graph. Fuzzy membership functions are used to map the overlap rate and remaining width to three levels of interference intensity: high, medium, and low. Adaptive pulse width modulation adjusts the duty cycle of the control signal according to the dynamic interference intensity level.

[0116] The dynamic interference intensity level is defined as follows:

[0117] ;

[0118] In the formula, Indicates the dynamic interference intensity level. Indicates a dynamic conflict identifier. This indicates a safe space to avoid.

[0119] It should be noted that the highest interference level is 3. When the dynamic conflict indicator detects a conflict and the avoidance space safety indicator fails, it means that the robot is simultaneously facing the risk of dynamic obstacle collision and the exhaustion of avoidance space, and the highest priority control response needs to be triggered.

[0120] The highest level of interference is 2. When only dynamic conflict markers detect a conflict or only avoidance space safety markers fail, it indicates the existence of a single risk source, requiring moderate-intensity adjustments.

[0121] The highest interference level is 1. If neither the dynamic conflict indicator nor the avoidance space safety indicator is triggered, it will operate in normal mode.

[0122] The physical feasibility of the control signal is verified through real-time simulation, invalid signals are filtered out and the timing signal is reconstructed;

[0123] Specifically, the dynamic behavior tree-generated motor control signal sequence is loaded in real-time simulation and compared with the obstacle geometric constraints and feasible region probability distribution in the environmental feature map. The dynamic feasibility of steering and braking pulses is verified by combining the Lyapunov stability function, and invalid signals exceeding the kinematic acceleration limit or conflicting with the obstacle distance field are filtered out. An incremental online distillation mechanism is used to align the valid signals with the constraints output by the quantum-inspired instruction recombination algorithm in a spatiotemporal tensor, and the phase of the timing signal is reconstructed through an adaptive pulse width modulation duty cycle smoothing mechanism.

[0124] The servo motor and brake are driven by the reconstructed timing signals to synchronously execute steering and braking actions.

[0125] Specifically, Lie group manifold operations are used to map the steering angle pulse duty cycle to the servo motor's angle increment, and the braking pulse amplitude to the brake's pressure gradient. Real-time feedback data from LiDAR and a vision camera is used to verify the kinematic consistency of the servo motor's angle and brake pressure in the environmental feature map. When a conflict is detected between the pulse signal and the probability distribution of the feasible region in the dynamic knowledge graph, a quantum-inspired instruction reordering algorithm is triggered to regenerate the pulse timing. An incremental online distillation mechanism is used to synchronously update the control parameters of the servo motor and brake, achieving synchronized execution of steering and braking actions.

[0126] The execution status is monitored in real time through an ultra-low latency feedback channel. If signal drift is detected, a self-repair protocol is triggered to regenerate the control signal.

[0127] Specifically, the state data of the servo motor angle encoder and brake pressure sensor are acquired in real time through an ultra-low latency feedback channel and input into the path connectivity rule base in the dynamic knowledge graph for kinematic consistency verification. When the deviation between the servo motor angle pulse duty cycle and the curvature derivative of the planned path exceeds the Lyapunov stability boundary in the dynamic knowledge graph, it is determined as a signal drift event, triggering a quantum-inspired instruction reconfiguration algorithm to reconstruct the superposition state of the control signal. The self-repair protocol fuses the phase of real-time sensor data and the reconstructed timing signal through an incremental online distillation mechanism, and regenerates the control signal using an adaptive pulse width modulation duty cycle smooth transition strategy.

[0128] S5. Collect the robot's actual path data and compare it point by point with the expected path. If the continuous deviation feature exceeds the preset continuous deviation feature threshold, input the environmental feature map, obstacle avoidance instruction set and continuous deviation feature into the deep reinforcement learning model for updating.

[0129] Furthermore, by fusing lidar and odometry for positioning, the robot's actual path data is collected in real time and compared point by point with the expected path to generate a path deviation sequence.

[0130] Specifically, by fusing LiDAR point cloud data with odometry encoder pulses for localization, and using Kalman filtering to predict the robot's real-time pose in the environmental feature map, the actual path coordinate sequence is extracted. The actual path coordinates are then compared with the expected path planning results in the dynamic knowledge graph using point-to-point Euclidean distance calculations to generate a path deviation sequence.

[0131] Perform dynamic trend analysis on the path deviation sequence to extract persistent deviation characteristics;

[0132] Specifically, the spatiotemporal evolution features of the path deviation sequence are extracted using a Kalman filter time series model. Spatiotemporal tensor decomposition is employed to separate the instantaneous fluctuation term and the persistent trend term of the path deviation sequence. Based on the path connectivity rule base in a dynamic knowledge graph, the criteria for determining persistent deviation features are defined. When the path deviation trend term within a continuous time window satisfies the instability criterion of the Lyapunov function, it is marked as a persistent deviation feature, triggering a quantum-inspired instruction reconfiguration algorithm to reconstruct the steering angle compensation parameters.

[0133] If the persistent deviation feature exceeds the preset persistent deviation feature threshold, the current environment feature map, obstacle avoidance instruction set and persistent deviation feature are associated to construct incremental training samples;

[0134] It should be noted that the preset persistent deviation feature threshold is based on the robot's maximum path tracking error tolerance calibration defined in the path connectivity rule base in the dynamic knowledge graph; for example, in the dynamic obstacle avoidance test scenario, the preset persistent deviation feature threshold is set to 0.5.

[0135] Specifically, the system retrieves spatiotemporal windows where persistent deviation features exceed a preset persistent deviation feature threshold from the path connectivity rule base in the dynamic knowledge graph. It then associates these windows with the obstacle geometry distribution, feasible region probability, and steering angle and braking level parameters from the obstacle avoidance command set in the corresponding timestamp environment feature map. Spatiotemporal tensor decomposition is used to extract a joint feature vector of environment-command-deviation. A quantum-inspired command reconfiguration algorithm maps these feature vectors to input-output pairs for incremental training samples. An incremental online distillation mechanism injects these samples into the policy distillation process of the deep reinforcement learning model, generating an incremental training sample set.

[0136] It should be noted that if the persistent deviation feature does not exceed the preset persistent deviation feature threshold, the path connectivity rule base in the dynamic knowledge graph determines that the current path tracking error is within the allowable tolerance, and the obstacle avoidance instruction set maintains the original steering angle and braking level parameters.

[0137] The parameters of the deep reinforcement learning model are dynamically updated based on incremental training samples.

[0138] Specifically, an incremental online distillation mechanism loads environmental feature maps, obstacle avoidance instruction sets, and persistent bias feature vectors from incremental training samples, and extracts a joint embedding representation of environment-instruction-bias through a neural symbolic co-encoding network. This embedding representation is then input into the policy network of the deep reinforcement learning model, where a fuzzy membership function in the differentiable symbolic logic layer is used to calculate the policy gradient direction. The gradient update range is then constrained by a path connectivity rule base within a dynamic knowledge graph.

[0139] This embodiment also provides a robot automatic obstacle avoidance target recognition system, including: an environment fusion module, a strategy generation module, a dynamic adjustment module, a control execution module, and a closed-loop optimization module; the environment fusion module is used to acquire and fuse obstacle boundary maps, passable area masks, and dynamic target heatmaps to output an environment feature map; the strategy generation module is used to input the environment feature map into a pre-trained deep reinforcement learning model and generate an obstacle avoidance instruction set based on the robot's real-time state and global path planning results; the dynamic adjustment module is used to predict the remaining width of the avoidance space, identify ground no-turn signs, predict the trajectory of dynamic targets, dynamically adjust the steering angle and braking level, and output a corrected obstacle avoidance instruction set based on the obstacle avoidance instruction set; the control execution module is used to convert the corrected obstacle avoidance instruction set into control signals to drive the robot to perform obstacle avoidance actions; the closed-loop optimization module is used to collect the robot's actual path data, compare it point by point with the expected path, and if the continuous deviation feature exceeds a preset continuous deviation feature threshold, input the environment feature map, obstacle avoidance instruction set, and continuous deviation feature into the deep reinforcement learning model for updating.

[0140] This embodiment also provides a computer device applicable to the robot automatic obstacle avoidance target recognition method, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to realize the robot automatic obstacle avoidance target recognition method proposed in the above embodiment.

[0141] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0142] This embodiment also provides a storage medium storing a computer program, which, when executed by a processor, implements the robot automatic obstacle avoidance target recognition method proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0143] In summary, this invention enhances the obstacle avoidance capability of autonomous robots in dynamic and complex environments through: multimodal perception fusion and closed-loop dynamic optimization mechanisms; constructing a spatiotemporally correlated global environmental feature representation through cross-modal fusion of multimodal data; utilizing differentiable modeling to analyze the continuous motion trend of dynamic targets and enhancing the trajectory prediction accuracy of sudden obstacles; combining the dynamic reward mechanism of deep reinforcement learning with physical constraint fusion strategies to achieve coordinated optimization of rapid response and motion smoothness; and autonomously adapting to unknown environmental disturbances through incremental learning driven by closed-loop feedback, forming a full-process obstacle avoidance architecture with autonomous evolution capabilities.

[0144] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for automatic obstacle avoidance and target recognition in robots, characterized in that: include, The process involves acquiring and fusing obstacle boundary maps, passable area masks, and dynamic target heatmaps to output an environmental feature map. The specific steps are as follows: Calculate the dynamic confidence weight of each sensor based on the real-time noise parameters of the multimodal sensor; Based on dynamic confidence weights, the obstacle boundary map, passable area mask, and dynamic target heat map are aligned to a unified topological coordinate system; The obstacle boundary map, passable area mask, and dynamic target heat map are fused to output a fused spatiotemporal feature map. The specific steps are as follows. The obstacle boundary map, passable area mask, and dynamic target heat map are mapped to the joint manifold space to generate spatiotemporally unified multimodal projection data; Spatiotemporal tensor decomposition is performed on multimodal projection data to separate the principal components of the motion trend of dynamic targets and the principal components of the structure of static environment; By selecting the dominant feature subspace, it is reconstructed into a fused spatiotemporal feature map; Verify the consistency of the fused spatiotemporal feature maps based on the preset map topology, and output the environmental feature map; The environmental feature map is input into a pre-trained deep reinforcement learning model. Based on the robot's real-time state and global path planning results, an obstacle avoidance instruction set is generated. The specific steps are as follows. By using a neural symbolic joint encoding network, environmental feature maps, robot real-time states, and global path planning results are aligned across modalities in time and space to generate dynamic knowledge-enhanced feature vectors. Embed a differentiable symbolic logic layer in a pre-trained deep reinforcement learning model to generate original obstacle avoidance instructions based on feature vectors enhanced by dynamic knowledge. A dynamic knowledge graph is constructed based on the original obstacle avoidance commands. Conflicts between commands and physical rules are detected through graph inference, triggering a quantum-inspired command reorganization algorithm to eliminate the conflicts. Specifically, spatial semantic labels of obstacle boundaries, passable areas, and dynamic target areas are extracted based on the semantic segmentation results of the original obstacle avoidance commands. These semantic labels are mapped to entity nodes and relational edges in the dynamic knowledge graph, constructing a graph topology that includes obstacle locations, path connectivity, and dynamic trajectory constraints. A graph inference engine is then used to retrieve conflicts between path planning commands and physical rules governing obstacle spatial distribution and dynamic target movement trends within the dynamic knowledge graph. An incremental online distillation mechanism is adopted to synchronously update the recombined instructions with real-time sensor feedback to generate an obstacle avoidance instruction set; Based on the obstacle avoidance instruction set, predict the remaining width of the avoidance space, identify ground no-turn signs, predict the dynamic target trajectory, dynamically adjust the steering angle and braking level, and output the corrected obstacle avoidance instruction set. The modified obstacle avoidance instruction set is converted into control signals to drive the robot to perform obstacle avoidance actions; Collect the robot's actual path data and compare it point by point with the expected path. If the continuous deviation feature exceeds the preset continuous deviation feature threshold, input the environmental feature map, obstacle avoidance instruction set and continuous deviation feature into the deep reinforcement learning model for updating.

2. The robot automatic obstacle avoidance target recognition method as described in claim 1, characterized in that: The specific steps for outputting the corrected obstacle avoidance command set are as follows. Based on the robot's real-time speed and turning angle, dynamically predict the remaining width of the avoidance space and generate a safety marker for the avoidance space. The system uses a visual camera to identify ground no-turn signs. If a ground no-turn sign is detected and the confidence level exceeds a preset confidence threshold, a ground constraint sign is generated. Kalman filtering is used to predict the trajectory of dynamic targets, calculate the overlap rate with the robot path, and generate dynamic conflict markers. Adjust the steering angle and braking level based on the combined status of obstacle avoidance space safety signs, ground constraint signs, and dynamic conflict signs; Adjust parameters based on actual obstacle avoidance path deviation, and output the corrected obstacle avoidance command set.

3. The robot automatic obstacle avoidance target recognition method as described in claim 2, characterized in that: The specific steps for driving the robot to perform obstacle avoidance actions are as follows. Based on the steering angle and braking level in the revised obstacle avoidance instruction set, execution priorities are assigned through a dynamic behavior tree to generate motor steering and braking pulse control signals. Based on dynamic conflict markers and clearance space safety markers, dynamic interference intensity levels are defined, and adaptive pulse width modulation is used to adjust the signal duty cycle. The physical feasibility of the control signal is verified through real-time simulation, invalid signals are filtered out and the timing signal is reconstructed; The servo motor and brake are driven by the reconstructed timing signals to synchronously execute steering and braking actions. The execution status is monitored in real time through an ultra-low latency feedback channel. If signal drift is detected, a self-repair protocol is triggered to regenerate the control signal.

4. The robot automatic obstacle avoidance target recognition method as described in claim 1, characterized in that: The input deep reinforcement learning model is updated, and the specific steps are as follows. The robot's actual path data is collected in real time by fusion positioning of lidar and odometry, and the path deviation sequence is generated by comparing it point by point with the expected path. Dynamic trend analysis is performed on the path deviation sequence to extract persistent deviation characteristics; If the persistent deviation feature exceeds the preset persistent deviation feature threshold, the current environment feature map, obstacle avoidance instruction set and persistent deviation feature are associated to construct incremental training samples; The parameters of the deep reinforcement learning model are dynamically updated based on incremental training samples.

5. A robot automatic obstacle avoidance target recognition system, based on the robot automatic obstacle avoidance target recognition method according to any one of claims 1 to 4, characterized in that: It includes an environment integration module, a strategy generation module, a dynamic adjustment module, a control execution module, and a closed-loop optimization module; The environment fusion module is used to acquire obstacle boundary map, passable area mask and dynamic target heat map and fuse them to output environment feature map; The strategy generation module is used to input the environmental feature map into a pre-trained deep reinforcement learning model and generate an obstacle avoidance instruction set based on the robot's real-time state and global path planning results. The dynamic adjustment module is used to predict the remaining width of the obstacle avoidance space based on the obstacle avoidance instruction set, identify ground no-turn signs, predict the trajectory of the dynamic target, dynamically adjust the steering angle and braking level, and output the corrected obstacle avoidance instruction set. The control execution module is used to convert the modified obstacle avoidance instruction set into control signals to drive the robot to perform obstacle avoidance actions; The closed-loop optimization module is used to collect the robot's actual path data and compare it point by point with the expected path. If the continuous deviation feature exceeds the preset continuous deviation feature threshold, the environmental feature map, obstacle avoidance instruction set and continuous deviation feature are input into the deep reinforcement learning model for updating.

6. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the robot automatic obstacle avoidance target recognition method according to any one of claims 1 to 4.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the robot automatic obstacle avoidance target recognition method according to any one of claims 1 to 4.

Images