Warehouse robot avoidance mode determination method

Abstract

This invention discloses a method for determining obstacle avoidance strategies for warehouse robots, relating to the field of artificial intelligence technology. The system, during operation, includes: fusing data from LiDAR, inertial measurement units, and visual sensors to construct a dynamic environment model with semantic labels; generating behavioral intent vectors based on task type, urgency, and historical interactions, and forming a local collaborative intent graph through distributed communication; inputting the above information into a lightweight decision network to output obstacle avoidance strategies such as deceleration, lateral movement, detour, or slight waiting; and using a time-elastic band algorithm to plan a smooth trajectory that satisfies dynamic constraints, monitoring deviations during execution, and replanning locally when necessary. This application significantly improves obstacle avoidance efficiency, trajectory smoothness, and system robustness through context awareness, intent collaboration, and adaptive safety mechanisms, supporting the efficient and stable operation of large-scale robot swarms in high-density warehouse environments.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, specifically to a method for determining the obstacle avoidance method of a warehouse robot. Background Technology

[0002] With the widespread application of intelligent warehousing systems in logistics and supply chain management, warehouse robots, as the core carriers for automated storage, retrieval, and handling, directly impact the throughput and stability of the entire warehousing operation through their operational efficiency and path safety. Modern high-density warehousing environments typically deploy a large number of robots working collaboratively, requiring them to possess high-precision positioning, dynamic obstacle avoidance, and efficient path planning capabilities. Especially in scenarios with dense multi-robot scheduling, robots need to perceive surrounding moving obstacles (such as other robots, temporary stacks of goods, or personnel) in real time with centimeter-level positioning accuracy and quickly decide on the optimal avoidance strategy. This places stringent requirements on the system's positioning reliability, the timeliness of environmental modeling, and the adaptability of its avoidance logic.

[0003] However, existing obstacle avoidance methods based on fixed rules or simple thresholds struggle to handle multi-target interactions in complex dynamic environments, often leading to path oscillations or deadlocks. Meanwhile, while high-precision positioning data (such as UWB or laser SLAM output) provides accurate location information, traditional obstacle avoidance mechanisms lack the ability to jointly model positioning uncertainty and motion trends, failing to effectively predict the future trajectories of neighboring robots. Furthermore, most solutions treat obstacle avoidance decisions and global path planning separately, relying solely on local sensor information for reactive adjustments, ignoring contextual semantics such as task priority, direction of travel, and historical behavior patterns. This results in decreased overall scheduling efficiency and even cascading congestion.

[0004] Therefore, there is an urgent need for a method to determine the obstacle avoidance mode of warehouse robots that integrates high-precision positioning information and multi-dimensional environmental semantics. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a method for determining the obstacle avoidance mode of a warehouse robot, thus solving the problems mentioned in the background section.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a method for determining the obstacle avoidance mode of a warehouse robot, comprising the following specific steps: Step 1: Obtain high-precision positioning data and environmental semantic information: By fusing multi-source perception data from LiDAR, inertial measurement unit and vision sensor, and combining with high-precision map, the high-precision positioning information of the warehouse robot in the global coordinate system is obtained in real time, and the position, speed and motion trend of surrounding dynamic obstacles are extracted simultaneously to construct a dynamic environment model containing spatial geometry and semantic labels. Step 2: Construct a multi-robot collaborative intent representation: Based on the task type, task urgency, current motion state and historical interaction trajectory of each warehouse robot, generate a behavioral intent vector for each robot, and exchange the intent vector among neighboring robots through a distributed communication mechanism to form a local group collaborative intent map; Step 3: Generate context-aware avoidance decisions: Input the dynamic environment model and the local group collaborative intent graph into a pre-trained lightweight decision network. The network outputs an avoidance strategy for the local machine based on the current traffic density, path conflict level, and task priority, including deceleration, lateral movement, detour, or waiting. Step 4: Plan a smooth avoidance trajectory. Based on the avoidance strategy, the time elastic band algorithm is used to generate a continuous and differentiable avoidance trajectory while satisfying the dynamic constraints. This ensures that the rate of change of trajectory curvature does not exceed a preset threshold, thus avoiding sudden stops or sharp turns. Step 5: Execute and provide feedback on the avoidance action: Control the warehouse robot to move along the avoidance trajectory, while continuously monitoring the deviation between the actual trajectory and the planned trajectory. If the deviation exceeds the allowable range, trigger the local replanning module to correct the trajectory without changing the original avoidance strategy until the avoidance action is completed.

[0007] Preferably, the acquisition of high-precision positioning data in step 1 adopts a multi-sensor fusion framework based on factor graph optimization, wherein the lidar point cloud matching frequency, the inertial measurement unit sampling frequency, and the visual feature extraction frame rate are all set to preset frequencies that meet real-time requirements, the positioning error is less than a preset accuracy threshold, and the update delay is less than a predetermined time threshold.

[0008] Preferably, the behavioral intent vector in step 2 is composed of four-dimensional parameters, which respectively represent the task type weight, the reciprocal of the remaining task time, the current speed vector magnitude, and the historical frequency of avoidance in the most recent interactions. Each parameter is normalized and then weighted and synthesized, and the weight coefficient is dynamically adjusted according to the functional zoning of the warehouse area.

[0009] Preferably, the lightweight decision network in step 3 is a multi-layer fully connected neural network. The number of nodes in the input layer, the number of nodes in the hidden layer, and the number of nodes in the output layer are set according to the task complexity. The output layer corresponds to multiple avoidance strategies. The network is trained in a large number of simulation scenarios through reinforcement learning in the offline stage, and the inference time during deployment is less than a preset time threshold.

[0010] Preferably, in step 4, the time resolution, trajectory point interval, and curvature change rate threshold of the time elastic band algorithm are all set as predetermined parameters to ensure stable operation and to ensure that the robot's maximum lateral acceleration does not exceed the preset acceleration threshold.

[0011] Preferably, in step 5, the allowable range of trajectory deviation is set such that both the lateral deviation and the longitudinal deviation are less than their respective preset tolerance thresholds. If the deviation exceeds this range for several consecutive control cycles, local replanning is initiated. The length of the replanning window and the number of trajectory points adjusted are set according to the stability requirements of the avoidance strategy to maintain the overall avoidance intention unchanged.

[0012] Preferably, the semantic labels in the dynamic environment model include several categories such as stationary shelves, mobile forklifts, pedestrians, other warehouse robots, and temporary obstacles. Each type of obstacle is assigned a different safe distance threshold. For other warehouse robots, an adaptive safe distance is adopted, which is dynamically adjusted as the relative speed increases, and its value is between a preset minimum and maximum value.

[0013] Preferably, the coverage radius of the local group collaborative intent graph is a predetermined distance, the nodes in the graph represent neighboring robots, and the edge weight is determined by the intersection angle of the two robot paths and the relative speed. When the intersection angle is less than a preset angle threshold and the relative speed is greater than a preset speed threshold, a high-priority collaborative negotiation mechanism is triggered.

[0014] Preferably, the "waiting" action in the avoidance strategy is not completely still, but rather involves slight swaying or rotation in place to maintain sensor data updates and avoid being misjudged as a static obstacle. The swaying amplitude is controlled within a preset angle range, and the frequency is a preset operating frequency.

[0015] Preferably, the method is deployed on edge computing nodes, each warehouse robot is equipped with an independent decision-making unit, and is connected to a regional coordinator. The regional coordinator is responsible for monitoring deadlock risks. When it detects that more than a predetermined number of robots are in a state of mutual waiting within a predetermined time period, it forcibly designates one of them to execute the priority passage instruction, thereby breaking the circular dependency.

[0016] This invention provides a method for determining the obstacle avoidance mode of a warehouse robot, which has the following beneficial effects: (1) During system operation, compared with the prior art, the collaborative obstacle avoidance method for warehouse robots provided by this invention combines multi-source perception fusion with intent-driven decision-making to achieve efficient and safe obstacle avoidance control in complex and high-density warehouse environments. Specifically, by fusing data from lidar, inertial measurement units, and visual sensors, and combining them with high-precision maps, this invention can obtain high-precision status information of the robot and surrounding dynamic obstacles in real time in a global coordinate system, thereby significantly improving the integrity and positioning accuracy of environmental perception and reducing the risk of misjudgment caused by perception uncertainty. Furthermore, by constructing a multi-robot collaborative intent representation, this invention uniformly encodes information such as task type, urgency, motion state, and historical interaction trajectory into a behavioral intent vector, and performs distributed exchange among neighboring robots, enabling each robot to have the ability to predict the behavioral trends of others within a local range, effectively reducing the problem of frequent start-stop or repeated conflicts in traditional passive obstacle avoidance mechanisms, and improving overall traffic efficiency. Furthermore, by inputting a dynamic environment model and a collaborative intent graph into a lightweight decision network, differentiated avoidance strategies can be adaptively generated based on traffic density, path conflict level, and task priority. This ensures that avoidance decisions balance safety and task efficiency, avoiding the resource waste caused by a "one-size-fits-all" approach to avoidance. The smooth avoidance trajectory generated by the time-elastic band algorithm satisfies dynamic constraints while limiting the rate of curvature change, guaranteeing the continuity and executability of robot motion and reducing mechanical wear and control jitter. Finally, a trajectory deviation monitoring and local replanning mechanism is introduced during the execution phase to correct the trajectory in real time without altering the predetermined avoidance strategy, enhancing the system's robustness to environmental disturbances and execution errors. In summary, this invention can improve the overall operating efficiency, stability, and scalability of multi-robot systems while ensuring safe avoidance, making it suitable for high-density, highly dynamic warehouse operation scenarios.

[0017] (2) By integrating multi-source sensor fusion and semantic modeling, high-precision positioning and dynamic obstacle intent recognition are unified. The environmental perception update frequency reaches above the preset high-frequency threshold, significantly improving perception integrity compared to the traditional single lidar solution and effectively eliminating collision risks caused by perception blind spots. By introducing behavioral intent vectors and local collaborative intent graphs, avoidance decision-making is transformed from passive response to active collaboration. The path conflict resolution time is shortened to within the preset time threshold, the system throughput is significantly improved in high-density scenarios, and the deadlock rate is reduced to an extremely low level.

[0018] (3) The time elastic band algorithm is used to generate a dynamically feasible avoidance trajectory. Combined with the curvature change rate constraint, the number of robot start-stop times is greatly reduced, the path oscillation amplitude is significantly reduced, the life of mechanical parts is significantly extended, and the smoothness of operation is improved. Through adaptive safety distance, micro-waiting action and edge deadlock intervention mechanism, the system can still maintain safe operation under abnormal conditions such as sudden obstacles, brief communication interruption or instantaneous sensor failure. The fault recovery time is less than the preset recovery time limit, and the reliability reaches the industrial application standard.

[0019] (4) The decision network reasoning time is lower than the preset time threshold, the single machine computing resources are less than the predetermined power consumption threshold, it can run on a low-cost embedded platform, support the deployment of a large number of robots in a single warehouse for collaborative operation, and has strong scalability to meet the needs of future large-scale development of intelligent warehousing. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the overall technical solution architecture of the method for determining the obstacle avoidance mode of a warehouse robot proposed in this invention; Figure 2 This is a schematic diagram of the core principle framework of multi-robot collaborative intent representation and situation-aware avoidance decision-making in this invention; Figure 3 This is a flowchart illustrating the logical flow of dynamic environment modeling, avoidance strategy generation, and smooth trajectory planning in this invention. Detailed Implementation

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. 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.

[0022] Example 1 This invention provides a method for determining the obstacle avoidance mode of a warehouse robot. Please refer to [link / reference]. Figures 1 to 3Step 1, acquiring high-precision positioning data and environmental semantic information, specifically includes the following operational procedures: First, the LiDAR deployed on the warehousing robot body collects three-dimensional point cloud data of the surrounding environment at a frequency of 10 Hz. After the point cloud data undergoes noise reduction, downsampling, and coordinate system alignment operations by the preprocessing module, it is input into a multi-sensor fusion framework based on factor graph optimization. The inertial measurement unit outputs raw data from the three-axis accelerometer and three-axis gyroscope at a sampling frequency of 200 Hz. After zero-bias correction and temperature compensation, state prediction is performed through an extended Kalman filter. Simultaneously, the wide-angle vision sensor installed on the top of the robot captures RGB images at a rate of 30 frames per second, uses the ORB feature extraction algorithm to generate scale-invariant feature point sets in real time, and calculates the camera pose using the PnP algorithm. After timestamp alignment, the data from the three types of sensors are used to construct a factor graph model containing pose nodes and constraint edges. The pose nodes represent the robot's six-DOF pose in the global coordinate system, and the constraint edges include pose constraints generated by LiDAR point cloud matching, kinematic constraints generated by inertial measurement unit integration, and observation constraints constituted by visual feature reprojection errors. This factor graph is incrementally optimized using the iSAM2 algorithm, outputting high-precision positioning results with an update frequency of no less than 20 times per second. Its positioning error is less than 0.02 meters within a 95% confidence interval, and the update latency is less than 50 milliseconds.

[0023] Based on high-precision positioning, the system simultaneously constructs a dynamic environment model. This model consists of a spatial geometry layer and a semantic labeling layer. The spatial geometry layer projects the LiDAR point cloud onto a global coordinate system and performs voxelization to form a 3D grid map with a resolution of 0.1m × 0.1m × 0.1m. Each voxel is labeled as occupied, idle, or unknown. The semantic labeling layer uses a deep learning semantic segmentation network to perform pixel-level classification of visual images, identifying five categories of objects: stationary shelves, moving forklifts, pedestrians, other warehouse robots, and temporary obstacles. The classification results are then back-projected into 3D space, assigning corresponding semantic labels to the voxels. For dynamic obstacles, the system continuously tracks their position changes over multiple frames, calculating their velocity vector and motion trend parameters. The velocity vector is obtained by fitting the position data of the most recent five time steps using the least squares method, and the motion trend parameters include the acceleration direction angle and trajectory curvature. Furthermore, the system sets differentiated safety distance thresholds for different types of obstacles: the safety distance for stationary shelves is fixed at 0.3 meters; the safety distance for mobile forklifts is set at 0.8 meters; the safety distance for pedestrians is set at 1.0 meter; the safety distance for temporary obstacles is set at 0.5 meters; and for other warehouse robots, an adaptive safety distance mechanism is adopted, where the distance d... adaptive Determined by the following formula: ; in, The preset minimum safe distance is set to 0.25 meters; The preset maximum safe distance is set to 0.6 meters; The relative velocity modulus between the two robots; The preset maximum relative speed threshold is set to 2.0 m / s. This adaptive mechanism ensures increased safety margin during high-speed approach and reduced redundancy during low-speed or stationary approaches, thereby improving channel utilization.

[0024] In the above-mentioned method for determining the avoidance mode of a warehouse robot, step 2, which constructs a multi-robot collaborative intent representation, is specifically executed as follows: Each warehouse robot maintains a behavioral intent vector: , This vector consists of parameters in four dimensions. The first dimension, w... t The first dimension represents the task type weight, assigned a weight value of 1.0, 0.8, or 0.5 depending on whether the current task is picking, moving, or charging; the second dimension r t It is the countdown to the remaining time of the task, that is T remain Let r be the number of seconds remaining from the current time until the task deadline. If the remaining time is less than 10 seconds, then r... t Clamped to 0.1; third dimension v c The magnitude of the current velocity vector, in meters per second; the fourth dimension, h. a The frequency of avoidance actions performed in the last 10 interactions with other robots is normalized to the [0,1] interval. After independent normalization, the above four parameters are weighted and synthesized to obtain the final behavioral intent vector, whose synthesis formula is: Where norm(·) represents the linear normalization function, mapping each parameter to [0,1]; weight coefficients The sum is 1, and its specific value is dynamically adjusted according to the functional zoning of the warehouse area where the robot is currently located—for example, in the high-frequency picking area, The task urgency weight is set to 0.4, while it drops to 0.2 in the cache.

[0025] After generating the behavioral intent vector, the robot exchanges this information with neighboring robots through a distributed communication mechanism. Communication uses a mesh network based on the IEEE 802.11s standard, with a broadcast period of 100 milliseconds and a communication coverage radius of 8 meters. Upon receiving the intent vector from a neighboring robot, the robot constructs a local group cooperative intent graph. This graph is stored as a directed graph, where each node represents a neighboring robot. Node attributes include its ID, position, velocity, and complete behavioral intent vector. Each edge connects two robot nodes that may experience path conflicts, and the edge weight $w_{ij}$ is determined by the intersection angle of the two robot paths. With relative velocity The decision was made jointly, and the calculation method is as follows: ; in, , The angle between the paths planned for the two robots at the nearest intersection point (value range [0, π]). When the intersection angle... Less than 30 degrees (i.e., π / 6 radians) and relative speed When the speed exceeds 1.0 m / s, the system determines that a high-risk conflict exists and triggers a high-priority collaborative negotiation mechanism: At this time, the two robots suspend their regular avoidance decisions and enter a dedicated negotiation channel, comparing the comprehensive scores of their respective behavioral intention vectors (i.e., The lower scorer (based on a scalar value) assumes the responsibility of yielding and sends a confirmation signal to the other party. This negotiation process has a time limit of 200 milliseconds; if the time limit is exceeded, the regional coordinator will intervene to arbitrate.

[0026] In the aforementioned method for determining the obstacle avoidance mode of a warehouse robot, step 3 generates a context-aware obstacle avoidance decision. Specifically, this is achieved by inputting the dynamic environment model generated in step 1 and the local group collaborative intent map constructed in step 2 into a pre-trained lightweight decision network. This network is a three-layer fully connected neural network. The input layer contains 128 nodes, encoding the distribution of environmental obstacles (64-dimensional), the motion state of dynamic obstacles (32-dimensional), and the intent vectors of neighboring robots (32-dimensional). The hidden layer contains 64 nodes, employing the ReLU activation function. The output layer contains 4 nodes, corresponding to four obstacle avoidance strategies: deceleration, lateral movement, detour, and waiting, and outputs the probability distribution using the Softmax function. The network is trained offline using a Deep Q-Network (DQN) reinforcement learning algorithm across over 100,000 simulation scenarios, covering complex conditions such as high-density intersections, narrow passages, and sudden obstacle intrusions. The training objective is to maximize the long-term task completion efficiency and minimize the weighted sum of obstacle avoidance energy consumption. During deployment, network inference time is strictly controlled within 10 milliseconds to meet real-time requirements.

[0027] The input data to the decision network undergoes standardized preprocessing. Obstacle information in the dynamic environment model is converted into a polar coordinate grid centered on the robot, with an angular resolution of 5 degrees and a distance resolution of 0.2 meters, comprising 72 sectors. Each sector records the distance, category, and relative speed of the nearest obstacle. The cooperative intent graph is aggregated into statistical features, including the number of neighboring robots, average intent score, and highest conflict weight. In the probability distribution of the network output, if the probability of a certain strategy exceeds 0.7, that strategy is directly adopted; if no strategy exceeds the threshold, the strategy with the highest probability is selected, and a confidence label is added for reference by the subsequent trajectory planning module. Specifically, when the output strategy is "wait," the system does not completely stop the robot but initiates a micro-oscillation mode: the robot chassis performs a 5-degree rotation to the left and right at a frequency of 0.5 Hz, with the oscillation amplitude strictly controlled within ±5 degrees, and the duration is the waiting time specified by the strategy. This action ensures that the LiDAR and vision sensors continuously acquire changing environmental data, avoiding pose drift of the SLAM system or misjudgment by other robots as static obstacles due to prolonged stillness.

[0028] In the aforementioned method for determining the obstacle avoidance method for a warehouse robot, step 4, planning a smooth obstacle avoidance trajectory, specifically involves the following steps: Based on the obstacle avoidance strategy output in step 3, the Timed Elastic Band (TEB) algorithm is invoked to generate the obstacle avoidance trajectory. The TEB algorithm represents the trajectory as a series of pose points associated with timestamps. ,in The timestamp of the k-th point, and the time interval between adjacent points. The timeout is set to 0.1 seconds. The algorithm optimization objective is to minimize the total trajectory length, the rate of change of curvature, and the proximity to obstacles, while satisfying robot dynamics constraints. Specific constraints include: maximum forward speed not exceeding 1.5 m / s, maximum lateral acceleration not exceeding 0.3 m / s², and the rate of change of trajectory curvature (i.e., the derivative)... ,in The absolute value of (where is the curvature and s is the arc length) does not exceed 0.5 meters. .

[0029] During optimization, the TEB algorithm treats obstacles in the dynamic environment model as time-varying ellipsoids, with their principal axis aligned with the obstacle's velocity direction and their size dynamically scaled according to a safety distance threshold. For the "side-shift" strategy, the algorithm searches for feasible paths within a 0.3-0.6 meter range to the left or right of the original path; for the "detour" strategy, the algorithm generates a semi-circular or C-shaped detour path centered on the obstacle, with the detour radius determined based on the obstacle's size and the safety distance; for the "deceleration" strategy, the algorithm only adjusts the timestamp t. kThe algorithm maintains the spatial path but extends the time spent traversing conflict zones. For the "wait" strategy, the algorithm generates a small-amplitude oscillating trajectory centered on the current position, with the amplitude consistent with the aforementioned micro-oscillation. The optimization solution employs Sequential Quadratic Programming (SQP), with an upper limit of 20 iterations and a single planning time not exceeding 30 milliseconds. The generated trajectory is a continuously differentiable spline curve, ensuring that the robot operates without sudden stops or sharp turns, and that the peak lateral acceleration remains below 0.25 m / s². .

[0030] In the aforementioned method for determining the obstacle avoidance mode of a warehouse robot, step 5 executes and provides feedback on the obstacle avoidance action. The specific execution mechanism is as follows: the robot's underlying motion controller receives the smooth obstacle avoidance trajectory generated in step 4, decomposes it into velocity commands for each joint, and drives the hub motors to perform the motion. Simultaneously, the system monitors the deviation between the actual trajectory and the planned trajectory at a frequency of 50 Hz. The deviation is calculated using a discrete approximation of the Fréchet distance, evaluating the lateral deviation separately. longitudinal deviation The allowable range for lateral deviation is set to ±0.05 meters, and the allowable range for longitudinal deviation is set to ±0.1 meters. If any deviation exceeds the corresponding tolerance threshold within three consecutive control cycles (i.e., 60 milliseconds), the local replanning module is triggered.

[0031] The local replanning module corrects the trajectory without altering the original avoidance strategy. For example, if the original strategy is "detour," but the actual trajectory deviates due to ground slippage, the replanning only adjusts trajectory points within a 0.2-meter radius of the original detour path, maintaining the detour intention. The replanning window length is set to 1.5 seconds, corresponding to 15 trajectory points. The number of trajectory points adjusted is dynamically determined based on the deviation: the larger the deviation, the more points are adjusted, up to a maximum of all 15 points. The replanning also uses the TEB algorithm, but the initial solution is an offset version of the original trajectory, resulting in faster convergence. After replanning, the new trajectory is immediately sent to the motion controller, and the old trajectory is discarded. This process is repeated until the avoidance action is completed, i.e., the robot passes through the conflict area and the distance to all obstacles is restored to a safe distance or higher.

[0032] To further verify the technical effects of this invention, the following specific application example is constructed: In a smart e-commerce warehouse center, 100 warehouse robots are deployed to perform order picking tasks. The warehouse layout includes a main aisle (2.5 meters wide), a secondary aisle (1.8 meters wide), and a dense shelving area (aisles 1.2 meters wide). At a certain moment, robot A (task type: emergency replenishment, 8 seconds remaining) travels north along the main aisle at a speed of 1.2 meters per second, while robot B (task type: routine handling, 120 seconds remaining) merges from the secondary aisle on the east side at a speed of 1.0 meter per second. The paths of the two robots form an angle of approximately 25 degrees at the intersection.

[0033] System execution step 1: LiDAR, IMU and visual sensor fusion positioning, positioning error 0.015 meters, update delay 40 milliseconds; the dynamic environment model identifies robot B as "other warehouse robot", relative speed 1.8 m / s, adaptive safety distance calculation is 0.25 + (1.8 / 2.0)×(0.6-0.25) = 0.565 meters.

[0034] Step 2: Robot A's behavioral intent vector is [1.0, 0.125, 1.2, 0.3], with a normalized overall score of 0.82; Robot B's vector is [0.8, 0.0083, 1.0, 0.6], with an overall score of 0.55. Because the intersection angle is 25 degrees < 30 degrees and the relative speed is > 1.0 m / s, high-priority negotiation is triggered, and Robot B, due to its lower score, assumes the responsibility of avoiding the obstacle.

[0035] Step 3: The decision network takes the environment and intent data as input and outputs the "side-shift" strategy with a probability of 0.85.

[0036] Step 4: The TEB algorithm generates a lateral trajectory 0.4 meters to the right of robot B's original path, with a maximum rate of curvature change of 0.42 meters. The peak lateral acceleration is 0.22 m / s². .

[0037] Step 5: Robot B performs a lateral movement. The actual trajectory has a maximum lateral deviation of 0.04 meters, which is within the tolerance. The avoidance is successful. The entire process takes 2.1 seconds and does not affect Robot A's emergency task.

[0038] Example 2 This embodiment is an explanation based on Embodiment 1. Please refer to it. Figure 1 Specifically, in another application scenario, consider the abnormal condition of a brief communication interruption. A warehouse robot C is traveling towards robot D in a narrow passage (1.3 meters wide). Both detect each other and begin to avoid each other. However, during the negotiation process, robot C's communication module is interrupted for 200 milliseconds due to electromagnetic interference.

[0039] The system ensures robustness through the following mechanisms: In step 2, the collaborative intent graph has a timeout failure mechanism. If no update is received from a neighboring robot within 200 milliseconds, its intent vector is set to the default value (task weight 0.5, remaining time countdown 0.01, speed 0.5, avoidance frequency 0.5). The decision network in step 3 already includes communication interruption scenarios during training, and therefore tends to choose a conservative strategy. In this example, the network outputs a "wait" strategy.

[0040] During step 5, robot C initiated a slight oscillation (±5 degrees, 0.5 Hz) to maintain sensor data updates. Simultaneously, the area coordinator detected that both robots were in a "waiting" state for 3 seconds, with a distance of less than 1.5 meters between them, indicating a potential deadlock risk. Since the system was set to a "predetermined quantity" of 2 robots and a "predetermined time period" of 3 seconds, the coordinator forcibly designated robot D to execute the priority passage command. Robot D switched to a "deceleration" strategy and slowly passed through, while robot C continued to wait until the passage was clear. The entire fault recovery process took 4.2 seconds without a collision.

[0041] Example 3 This embodiment is an explanation based on Embodiment 1. Please refer to it. Figure 1 Specifically: In high-density operation scenarios, 20 robots converge simultaneously at the sorting area exit.

[0042] The system improves throughput in the following ways: In step 1, the dynamic environment model marks other robots as high-priority obstacles and dynamically adjusts the adaptive safety distance; In step 2, the collaborative intent graph coverage radius is expanded to 10 meters, incorporating more interactive nodes; Because of the high input traffic density, the decision network in step 3 tends to output a combination of "detour" and "lateral movement" to reduce the risk of complete stoppage. The TEB algorithm in step 4 employs parallel optimization, with multiple avoidance trajectories generated simultaneously; The local replanning frequency in step 5 is increased to once every two control cycles to ensure that the trajectory closely matches the dynamic environment.

[0043] Real-world testing shows that the system resolves path conflicts in this scenario in an average of 1.8 seconds, with a deadlock rate of 0.02%, which is significantly better than traditional methods.

[0044] Example 4 This embodiment is an explanation based on Embodiment 1. Please refer to it. Figure 1 Specifically, for momentary sensor failures, such as the loss of one frame of data from a LiDAR sensor due to dust obstruction, the system ensures positioning continuity through multi-source redundancy. In the factor graph optimization framework of step 1, the inertial measurement unit provides short-term, high-frequency motion prediction, and the visual sensor supplements geometric features. Even with missing LiDAR data, the positioning error can still be controlled within 0.05 meters. The dynamic environment model extrapolates obstacle positions using historical trajectories to maintain the effectiveness of obstacle avoidance decisions. After the sensors recover, the system quickly re-initializes the synchronization state to avoid abrupt trajectory changes.

[0045] Example 5 A method for determining obstacle avoidance strategies for warehouse robots, please refer to... Figures 1 to 3Specifically, in edge computing deployment mode, each robot is equipped with an NVIDIA Jetson AGX Orin embedded platform, running an independent decision-making unit. The lightweight decision network is optimized with TensorRT, achieving an inference time of 8 milliseconds and a power consumption of 15 watts. The regional coordinator is deployed on a local server, accessed via gigabit Ethernet, and monitors the status of all robots in the warehouse. When more than 5 robots are detected waiting for each other within 5 seconds, deadlock intervention is initiated. This architecture supports the deployment of 500 robots in a single warehouse, offering strong scalability and meeting the needs of large-scale cluster applications.

[0046] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for determining the obstacle avoidance mode of a warehouse robot, characterized in that: Includes the following steps: By acquiring high-precision positioning data and environmental semantic information, and by fusing multi-source perception data from LiDAR, inertial measurement unit and vision sensor, combined with high-precision map, the high-precision positioning information of the warehouse robot in the global coordinate system is obtained in real time. At the same time, the position, speed and motion trend of surrounding dynamic obstacles are extracted to construct a dynamic environment model containing spatial geometry and semantic labels. Construct a multi-robot collaborative intent representation. Based on the task type, task urgency, current motion state and historical interaction trajectory of each warehouse robot, generate a behavioral intent vector for each robot, and exchange the intent vector among neighboring robots through a distributed communication mechanism to form a local group collaborative intent map. A context-aware avoidance decision is generated by inputting the dynamic environment model and the local group collaborative intent graph into a pre-trained lightweight decision network. The network outputs an avoidance strategy for itself based on the current traffic density, path conflict level, and task priority. The avoidance strategy includes deceleration, lateral movement, detour, or waiting. A smooth avoidance trajectory is planned. Based on the avoidance strategy, a time elastic band algorithm is used to generate a continuously differentiable avoidance trajectory while satisfying dynamic constraints, ensuring that the trajectory curvature change rate does not exceed a preset threshold. The system executes and provides feedback on the avoidance action, controlling the warehouse robot to move along the avoidance trajectory. At the same time, it continuously monitors the deviation between the actual trajectory and the planned trajectory. If the deviation exceeds the allowable range, it triggers the local replanning module to correct the trajectory without changing the original avoidance strategy until the avoidance action is completed.

2. The method for determining the obstacle avoidance mode of a warehouse robot according to claim 1, characterized in that: The acquisition of high-precision positioning data adopts a multi-sensor fusion framework based on factor graph optimization. LiDAR point cloud matching, inertial measurement unit sampling, and visual feature extraction are all performed at a frequency that meets real-time requirements. The positioning error is less than a preset accuracy threshold, and the update delay is less than a predetermined time threshold.

3. The method for determining the obstacle avoidance mode of a warehouse robot according to claim 1, characterized in that: The behavioral intent vector consists of four parameters, representing the task type weight, the reciprocal of the remaining task time, the current speed vector magnitude, and the frequency of avoidance in the most recent interactions. Each parameter is normalized and then weighted and synthesized, and the weight coefficients are dynamically adjusted according to the functional zoning of the warehouse area.

4. The method for determining the obstacle avoidance mode of a warehouse robot according to claim 1, characterized in that: The lightweight decision network is a multi-layer fully connected neural network. The number of nodes in the input layer, hidden layer, and output layer is set according to the task complexity. The output layer corresponds to multiple avoidance strategies. The network is trained in a large number of simulation scenarios through reinforcement learning in the offline stage. When deployed, the inference time is less than the preset time threshold.

5. The method for determining the obstacle avoidance mode of a warehouse robot according to claim 1, characterized in that: The time resolution, trajectory point interval, and curvature change rate threshold of the time elastic band algorithm are all set as predetermined parameters to ensure stable operation and to ensure that the robot's maximum lateral acceleration does not exceed the preset acceleration threshold.

6. The method for determining the obstacle avoidance mode of a warehouse robot according to claim 1, characterized in that: The allowable range of trajectory deviation is set so that both the lateral deviation and the longitudinal deviation are less than their respective preset tolerance thresholds. If the deviation exceeds this range for several consecutive control cycles, local replanning is initiated. The length of the replanning window and the number of trajectory points adjusted are set according to the stability requirements of the avoidance strategy to keep the overall avoidance intention unchanged.

7. The method for determining the obstacle avoidance mode of a warehouse robot according to claim 1, characterized in that: The semantic labels in the dynamic environment model include stationary shelves, mobile forklifts, pedestrians, other warehouse robots, and temporary obstacles. Each type of obstacle is assigned a different safe distance threshold. For other warehouse robots, an adaptive safe distance is adopted, which is dynamically adjusted as the relative speed increases, and its value is between a preset minimum and a maximum value.

8. The method for determining the obstacle avoidance mode of a warehouse robot according to claim 1, characterized in that: The local group collaborative intent graph has a coverage radius of a predetermined distance. Nodes in the graph represent neighboring robots. The edge weights are determined by the intersection angle of the two robots' paths and their relative speeds. When the intersection angle is less than a preset angle threshold and the relative speed is greater than a preset speed threshold, a high-priority collaborative negotiation mechanism is triggered.

9. The method for determining the obstacle avoidance mode of a warehouse robot according to claim 1, characterized in that: The "waiting" action in the avoidance strategy is not completely still, but rather involves slight swaying or rotation in place to maintain sensor data updates and avoid being misjudged as a static obstacle. The swaying amplitude is controlled within a preset angle range, and the frequency is a preset operating frequency.

10. The method for determining the obstacle avoidance mode of a warehouse robot according to claim 1, characterized in that: The method is deployed on edge computing nodes. Each warehouse robot is equipped with an independent decision-making unit and is connected to a regional coordinator. The regional coordinator is responsible for monitoring deadlock risks. When it detects that more than a predetermined number of robots are stuck in a mutual waiting state within a predetermined time period, it forcibly designates one of them to execute the priority passage instruction, thus breaking the circular dependency.

Images