A dynamic warehouse navigation method and system for decoupling task logic and navigation strategy for single-wave picking

Abstract

The application provides a dynamic warehouse navigation method and system for single-wave sorting task logic and navigation strategy decoupling, adopts a high-fidelity, strong-random, noise-containing simulation environment as a test framework; adopts a deep reinforcement learning single-target navigation network ColorDynamic pre-trained by an independent bottom motion control layer as a bottom motion control to realize a scheduling strategy between single targets and relative dynamic obstacles; adopts an independent upper task management layer to form an order task queue; in a robot navigation process, the robot is regulated and controlled in motion behavior based on the bottom motion control layer, and continuous execution of tasks is performed based on a multi-target sequence of the upper task management layer, so that seamless continuous navigation of multiple targets is completed.

Description

Technical Field

[0001] This invention relates to the field of intelligent logistics and autonomous navigation technology for mobile robots. Specifically, it relates to a dynamic warehouse navigation method and system that achieves high robustness for single-wave picking tasks by decoupling task timing logic from the underlying navigation strategy in environments with strong dynamic interference, partial observability, and noise. Background Technology

[0002] With the rapid development of e-commerce logistics, warehouse robots (AGVs / AMRs) are increasingly widely used in single-wave, multi-target picking tasks. Existing warehouse scheduling and navigation methods are mainly divided into traditional analysis methods, modeling and simulation methods, intelligent optimization methods, and hybrid optimization methods. However, when facing the frequently changing dynamic disturbances in real-world warehouse environments, existing technologies have significant shortcomings:

[0003] Insufficient dynamic response capability: Traditional methods and some modeling and simulation methods rely on static environment assumptions or globally accurate information, making it difficult to adapt to unpredictable dynamic obstacles such as pedestrian flow and forklifts in real-world scenarios. Existing research often assumes that obstacles move in a straight line at a constant speed during simulation verification, ignoring the nonlinear characteristics of objects in the real environment, such as frequent changes in speed and direction.

[0004] The contradiction between real-time performance and computational overhead: Although intelligent optimization methods such as genetic algorithms and particle swarm optimization have global search capabilities, they have high computational overhead and slow convergence, making it difficult to meet the requirements of high-frequency real-time control, and they are also slow to respond to sudden disturbances during operation.

[0005] Long-sequence task policies become increasingly complex: Existing end-to-end deep reinforcement learning navigation methods often hard-encode multi-stage task sequences into the policy network. As the number of task objectives increases, the state space of the policy network explodes, leading to training difficulties and poor generalization. Once faced with trajectory deviations caused by dynamic obstacles, it is prone to getting stuck in local optima or needing to replan, lacking a flexible task switching mechanism.

[0006] Insufficient fidelity in warehouse environment simulation: Although general simulation platforms have various disturbance generation capabilities, idealized environment settings are commonly used in current warehouse navigation research and practice to simplify experiments.

[0007] The dynamic models are simplistic: they often use obstacles with fixed initial positions and uniform linear motion, lacking systematic simulation of the random clustering of obstacle positions in real warehouses and discrete decision-making behaviors such as acceleration, deceleration, and steering. The specific distribution and trajectory of obstacles are also too regular, failing to cover the diversity of real-world scenarios.

[0008] Limitations of Noise Models: In simulating sensor errors, existing research generally tends to use mathematically convenient Gaussian noise models. However, real sensors are limited by hardware characteristics, and their measurement errors often include various non-ideal features such as quantization errors and truncation errors. Existing warehouse navigation research has failed to fully cover the diversity of this noise distribution, resulting in trained algorithms exhibiting insufficient adaptability when facing the complex and diverse perceptual uncertainties in the real world.

[0009] Fixed task scenarios: Existing tests are often limited to a single map layout or a fixed sequence of target points, lacking a systematic evaluation of multiple map layouts and random combinations of multiple target points.

[0010] Therefore, there is an urgent need for a navigation method that can effectively decouple complex task sequence management from underlying dynamic obstacle avoidance, and can systematically construct a high-fidelity dynamic environment in warehousing scenarios, including the introduction of multi-dimensional random perturbations of position, velocity, and acceleration, as well as diverse noise models, in order to improve the robustness and generalization ability of warehousing robots in real complex environments. Summary of the Invention

[0011] The present invention aims to overcome the above-mentioned defects and provide a dynamic warehouse navigation method and system that decouples the task logic and navigation strategy for single-wave picking.

[0012] This invention constructs an independent task logic management layer and a black-box underlying navigation execution unit, and utilizes the Adaptive Goal Switching (AGS) mechanism to achieve seamless state injection and continuous control between multiple subtasks without interrupting the operation of the underlying strategy.

[0013] Specifically, this invention addresses the problem of overly idealized environmental models in existing warehouse navigation research by proposing a high-fidelity environment construction strategy. Under a given number of dynamic obstacles, a multi-dimensional random perturbation model of position, velocity, and acceleration, along with a uniformly distributed observation noise model, is systematically introduced. This aims to enhance the realism of the environment, particularly the diversity of motion behavior and perceived noise, thereby improving the system's robustness and task completion rate under partially observable conditions. The specific scheme is as follows:

[0014] This invention provides a dynamic warehouse navigation method that decouples task logic and navigation strategy for single-wave picking, characterized by:

[0015] A high-fidelity, highly random, and noisy simulation environment is used as the test framework;

[0016] A deep reinforcement learning single-target navigation network, ColorDynamic, pre-trained with an independent underlying motion control layer, is used for underlying motion control to realize a scheduling strategy between a single target and dynamic obstacles.

[0017] An independent upper-level task management layer is used to form an order task queue;

[0018] During robot navigation, the robot's motion behavior is regulated by the underlying motion control layer, while the tasks are executed continuously according to the multi-objective sequence of the upper task management layer (i.e., tasks are issued all at once at the beginning of each round, forming a task queue, and the robot executes tasks continuously according to the task queue), thereby completing seamless continuous navigation of multiple objectives.

[0019] Furthermore, the present invention provides a dynamic warehouse navigation method for decoupling task logic and navigation strategy for single-wave picking, characterized by comprising the following specific steps:

[0020] Step S1. Construct a high-fidelity dynamic warehouse simulation environment;

[0021] Step S2. Load the pre-trained underlying navigation policy model;

[0022] Step S3. Initialize the independent task logic management layer;

[0023] Step S4. Perform closed-loop control and seamless switching to adaptive target.

[0024] Furthermore, the present invention provides a dynamic warehouse navigation method for decoupling task logic and navigation strategy for single-wave picking, characterized in that: in S1, the specific method for constructing a high-fidelity dynamic warehouse simulation environment is as follows:

[0025] S1-1. Configure the geometric parameters and gait parameters of the warehouse map;

[0026] S1-2. Initialize a preset number of dynamic obstacle sets in the environment and execute a multi-dimensional randomized perturbation strategy for the warehouse scenario.

[0027] Furthermore, the present invention provides a dynamic warehouse navigation method for decoupling task logic and navigation strategy for single-wave picking, characterized in that:

[0028] The above-mentioned multidimensional randomization perturbation strategies include:

[0029] Initial position randomization: The initial positions of dynamic obstacles are uniformly and randomly distributed within the effective area of ​​the map;

[0030] Initial velocity randomization: the initial velocity vector of the obstacle (v) x , v y Random sampling is performed within a selected range;

[0031] Randomization of motion behavior: During the motion, the obstacle performs a velocity update once in a selected gait, and the velocity change Δv is randomly selected within a selected range;

[0032] Geometric size randomization: The physical size of obstacles changes randomly between rounds;

[0033] Diverse observation noise injection: Superimpose random noise with a uniform distribution onto the robot sensor observation data.

[0034] Furthermore, the present invention provides a dynamic warehouse navigation method for decoupling task logic and navigation strategy for single-wave picking, characterized in that:

[0035] The aforementioned random noise includes: position measurement noise, direction measurement noise, linear velocity noise, angular velocity noise, and radar measurement noise.

[0036] Furthermore, the present invention provides a dynamic warehouse navigation method for decoupling task logic and navigation strategy for single-wave picking, characterized in that:

[0037] In S2, the specific method for loading the pre-trained underlying navigation policy model is as follows:

[0038] S2-1. Load an end-to-end navigation policy network ColorDynamic, which is pre-trained based on deep reinforcement learning, and encapsulate it as a black-box execution unit;

[0039] S2-2. Configure the input vector of the underlying navigation strategy model;

[0040] The input vector of the underlying navigation strategy model configured above is a noisy sequence of robot states over a specified period of time in the past;

[0041] The above state sequence refers to the set of the robot's relative geometric relationship, its own motion state, and the results of environmental perception at each moment.

[0042] Furthermore, the present invention provides a dynamic warehouse navigation method for decoupling task logic and navigation strategy for single-wave picking, characterized in that:

[0043] The aforementioned underlying navigation strategy model outputs discrete action selection or continuous action parameters based on the input vector;

[0044] The aforementioned underlying motion control layer, based on a dynamic model, transforms the selection of discrete actions or the parameters of continuous actions into linear and angular velocity control signals, thereby regulating the robot's motion behavior.

[0045] Furthermore, the present invention provides a dynamic warehouse navigation method for decoupling task logic and navigation strategy for single-wave picking, characterized in that:

[0046] In S3, the specific method for adding an independent task logic management layer is as follows: construct a task logic management module that is independent of the underlying strategy and is responsible for the task sequence management throughout the entire lifecycle;

[0047] Among them, including

[0048] Task queue construction: Based on a single wave of picking orders, randomly select K shelf locations from N map layouts as task targets, and generate a double-ended queue Q containing multiple picking point coordinates and destination coordinates to simulate the order-driven picking path in an actual warehouse.

[0049] Furthermore, the present invention provides a dynamic warehouse navigation method for decoupling task logic and navigation strategy for single-wave picking, characterized in that:

[0050] In S4, the specific method for seamlessly switching between closed-loop control and adaptive objectives is as follows: within each control cycle, the underlying strategy reasoning and the upper-level logic monitoring are executed synchronously.

[0051] Among them, including

[0052] State awareness and reasoning: Acquire current noisy observation data and calculate the current target G. curent The relative pose is input into the underlying policy model to obtain motion commands and execute them;

[0053] Arrival determination: Real-time calculation of the robot's current position and G curent Euclidean distance d t If d t If the value is less than the preset threshold ε, the current subtask is considered complete.

[0054] Adaptive State Injection: If the current subtask is completed and the number of remaining targets in queue Q is greater than 1, perform a dequeue operation to remove G. curent The new first element of the queue is read as the next target, and the target relative pose parameters in the input vector are instantly updated to the parameters corresponding to the next target. At this time, the underlying strategy, based on the updated target information, directly outputs a smooth motion command pointing to the next target in the next control cycle, realizing the continuous transition between subtasks.

[0055] Termination judgment: If the queue is cleared and the destination is reached, the task is considered successful; if a collision occurs or the number of steps exceeds the maximum threshold, the task is considered to have failed.

[0056] Furthermore, this invention also provides a dynamic warehouse navigation system that decouples task logic and navigation strategy for single-wave picking, characterized by: including...

[0057] The high-fidelity environment construction module executes multi-dimensional randomized perturbation strategies and diverse noise injections for warehousing scenarios.

[0058] The strategy execution module carries the pre-trained black-box navigation model;

[0059] The logic control module executes queue management and adaptive target switching logic seamlessly.

[0060] The data acquisition and analysis module calculates task success rate, collision rate, timeout rate, and average number of successful steps.

[0061] The function and effects of this invention:

[0062] The decoupling architecture of this invention enhances robustness: by separating complex task sequence logic from the neural network, the overcomplexity of the policy network is avoided. The underlying policy can focus on the core capability of dynamic obstacle avoidance, enabling the system to maintain high performance without retraining when facing unpredictable dynamic disturbances.

[0063] This invention achieves true continuous navigation: through the AGS mechanism, it realizes instantaneous injection and seamless switching of target state.

[0064] The high-fidelity environment construction strategy of this invention significantly improves generalization ability.

[0065] This invention addresses the characteristics of warehouse scenarios by systematically introducing comprehensive randomization of position, velocity, and acceleration into the simulation environment, based on a fixed number of obstacles. Even with a constant number of obstacles, the random distribution of their initial positions and their discrete acceleration and direction-changing behaviors during movement construct a highly complex dynamic scenario. This breaks through the limitations of simplified dynamic models commonly used in existing research, enabling highly robust dynamic warehouse navigation for single-wave picking tasks.

[0066] This invention introduces a uniformly distributed noise model to replace or supplement single Gaussian noise, which more accurately covers the quantization error characteristics of the sensor and enriches the diversity of noise distribution in the training data.

[0067] The present invention's solution, under both distributed and centralized simple layouts, was validated through statistical testing across 100 episodes. The results show that, with a task scale of 3, the system achieves a stable task success rate exceeding 90%, a collision rate below 10%, and a timeout rate below 2%. This demonstrates that the SW-DWNS system can effectively cope with strong dynamic interference and diverse sensing uncertainties.

[0068] The present invention has strong generalization ability in task scenarios: by randomly selecting K+1 targets from predefined shelves under three map layouts, it simulates the diverse order combinations in a real warehouse, effectively testing and proving the algorithm's generalization ability under different path planning challenges. Attached Figure Description

[0069] Figure 1 The overall system architecture diagram of the method of this invention shows the decoupling relationship and data flow between the task logic management layer and the underlying navigation execution unit.

[0070] Figure 2 : Schematic diagram of the robot's continuous motion trajectory. Detailed Implementation

[0071] 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 skilled in the art without creative effort are within the scope of protection of the present invention.

[0072] In this embodiment, a dynamic warehouse navigation method is provided that decouples the task logic and navigation strategy for single-wave picking, including the following steps:

[0073] Step S1: Construct a high-fidelity dynamic warehouse simulation environment;

[0074] Configure the geometric parameters and gait parameters of the warehouse map, such as setting the physical size of the map to 800×800 cm, the control frequency to 10Hz (0.1 seconds / step), and the maximum number of steps per round to 2000 steps. Depending on the actual application needs, parameters such as the physical size of the map and the control frequency can be set according to the size of the actual construction scene and the performance of the robot itself.

[0075] Initialize a preset set of dynamic obstacles in the environment and execute a multi-dimensional randomized perturbation strategy for the warehouse scenario:

[0076] Initial position randomization: The initial positions of dynamic obstacles are uniformly and randomly distributed within the effective map area; for example, the initial positions of dynamic obstacles are uniformly and randomly distributed within the effective map area of ​​[40, 760] × [40, 760] cm. This range excludes the 40 cm area at the map edge, ensuring that obstacles do not initially coincide with the boundary. This is intended to simulate the characteristic that AGVs and pedestrians typically do not cluster at the warehouse edge, ensuring a high degree of uncertainty in the spatial distribution of obstacles in each evaluation. Depending on the needs of the actual application, the initial positions of these dynamic obstacles can be adjusted as needed and according to the actual constructed spatial volume; similarly, the edge can be 0 or other values.

[0077] Initial velocity randomization: the initial velocity vector of the obstacle (v) x , v y Random sampling is performed within a selected range; for example, the initial velocity vector (v) of an obstacle. x , v y Random sampling within the range of [-20, 20] cm / s simulates the typical speed characteristics of human walking and low-speed movement of small AGVs in a warehouse environment, breaking the assumption of initial stationary or uniform speed. Similarly, the variable range of this initial speed can be adjusted according to robot performance, etc.

[0078] Randomization of motion behavior: During the motion, the obstacle performs a speed update once in a selected gait, and the speed change Δv is randomly selected within a selected range; for example, during the motion, the obstacle performs a speed update once every 0.1 seconds, and the speed change Δv is randomly selected from {-1, 0, 1} cm / 0.1 sec (i.e., the acceleration changes discretely within the range of [-10, 10] cm / sec²), simulating the discrete decision-making steering and speed change behavior of real pedestrians or vehicles, avoiding the idealized assumption of uniform linear motion;

[0079] Geometric size randomization: The physical size of obstacles changes randomly between rounds to simulate pedestrians or vehicles of different sizes;

[0080] Diverse observation noise injection: Uniformly distributed random noise is superimposed onto robot sensor observation data to simulate the quantization error characteristics of specific sensor types, supplementing the shortcomings of traditional Gaussian noise models. Specifically, this includes:

[0081] Position measurement noise: uniformly distributed within ±2 cm;

[0082] Orientation measurement noise: uniformly distributed within ±3.6° (±π / 50 radians);

[0083] Linear velocity noise: uniformly distributed within ±1 cm / s;

[0084] Angular velocity noise: uniformly distributed within the range of ±3.6° / second;

[0085] Radar measurement noise: uniformly distributed within ±1 cm.

[0086] Step S2: Load the pre-trained underlying navigation policy model;

[0087] Load an end-to-end navigation policy network ColorDynamic, which is pre-trained based on deep reinforcement learning, and encapsulate it as a black-box execution unit;

[0088] The input vector of the aforementioned underlying navigation strategy model includes a noisy robot state sequence over a specified past period. This state sequence refers to the set of the robot's relative geometric relationships (e.g., relative distance D2T and relative angle alpha to the current target point; note that the "current target point" is a dynamically switched node in the task queue), the robot's own motion state (e.g., current linear velocity vlinear and angular velocity vangular), and environmental perception LiDAR data (e.g., grouped noisy LiDAR scan results) at each moment (a custom time point, such as a frame). The underlying navigation strategy model outputs discrete action selections (e.g., forward movement, turning) or parameters for continuous actions based on the input vector. Subsequently, the underlying motion control layer, based on a dynamics model, transforms these discrete action selections or continuous action parameters into linear and angular velocity control signals, thereby regulating the robot's motion behavior.

[0089] Key features: This underlying model does not contain multi-target sequence memory, but focuses only on single-target tracking and dynamic obstacle avoidance.

[0090] Step S3: Initialize the independent task logic management layer;

[0091] Build a task logic management module independent of the underlying strategy, responsible for the entire lifecycle management of task sequences:

[0092] Task queue construction: Based on a single wave of picking orders, K shelf locations are randomly selected from three map layouts as task targets, and a double-ended queue (Deque) Q containing multiple picking point coordinates and destination coordinates is generated to simulate the order-driven picking path in the actual warehouse.

[0093] Step S4: Perform seamless switching between closed-loop control and adaptive target (AGS)

[0094] Within each control cycle, low-level strategy reasoning and high-level logic monitoring are executed synchronously:

[0095] State awareness and reasoning: Acquire current noisy observation data and calculate the current target G.curent The relative pose is input into the underlying policy model to obtain motion commands and execute them;

[0096] Arrival determination: Real-time calculation of the robot's current position and G curent Euclidean distance d t If d t If the value is less than the preset threshold ε, the current subtask is considered complete.

[0097] Adaptive State Injection: If the current subtask is completed and the number of remaining targets in queue Q is greater than 1: perform a dequeue operation to remove G. curent Read the new first element of the queue as the next target; without resetting the environment state, interrupting the operation of the underlying policy, or reinitializing the neural network weights, directly update the target relative pose parameters in the input vector to the parameters corresponding to the next target; based on the updated target information, the underlying policy directly outputs a smooth motion command pointing to the next target in the next control cycle, realizing a continuous transition between subtasks;

[0098] Termination judgment: If the queue is cleared and the destination is reached, the task is considered successful; if a collision occurs or the number of steps exceeds the maximum threshold, the task is considered to have failed.

[0099] Application experiments are as follows:

[0100] Example 1: Construction and High Dynamic Environment Verification of the SW-DWNS System

[0101] This embodiment constructs a dynamic warehouse navigation system (SW-DWNS) for single-wave picking tasks and verifies it in a high-fidelity simulation environment.

[0102] 1. Experimental Environment Configuration

[0103] Map specifications: Three warehouse maps with different layouts are set up, each with a physical size of 800×800 cm.

[0104] Control parameters: The robot control frequency is set to 10Hz, which means the control interval is 0.1 seconds / step; the maximum number of steps per episode is set to 2000 steps.

[0105] Multidimensional dynamic perturbation settings:

[0106] Obstacle quantity setting: In each evaluation scenario, a fixed number of dynamic obstacles is set to 6.

[0107] Initial position random: The initial position of dynamic obstacles is uniformly and randomly distributed within the range of [40, 760] × [40, 760] cm, excluding the edge 40 cm area to ensure that they do not coincide with the boundary, simulating the uncertainty of the distribution of the real work area.

[0108] Initial velocity random: The initial velocity of obstacles is randomly sampled within the range of [-20, 20] cm / s to simulate the low-speed movement characteristics of people and vehicles in the warehouse, and to avoid all obstacles starting from a standstill or at the same speed.

[0109] Random motion behavior: The change in obstacle velocity every 0.1 seconds is discretely and randomly selected from {-1, 0, 1} cm / 0.1 seconds (i.e., acceleration ∈ [-10, 10] cm / s²) to simulate natural turning and speed change. This is the core means of simulating real dynamic interference in this embodiment.

[0110] Random size: The size of obstacles changes randomly between rounds.

[0111] Diverse noise injection: Superimposing uniformly distributed noise onto sensor observations.

[0112] Location noise: ±2 cm;

[0113] Directional noise: ±3.6°;

[0114] Linear velocity noise: ±1 cm / s;

[0115] Angular velocity noise: ±3.6° / second;

[0116] Radar noise: ±1 cm.

[0117] All noises are distributed with equal probability within a specified interval, aiming to simulate the quantization error characteristics that may exist in the sensor, enrich the types of noise distribution in the environment, and improve the robustness of the algorithm to non-Gaussian noise.

[0118] Table 1 Core Parameter Configuration Table

[0119]

[0120] 2. Task Scenario Setting

[0121] Target point randomization: Randomly select 1-5 target points from three predefined map layout shelves as target points for a single task, simulating the order scale of 1-5 target points in a real "goods-to-person" warehouse.

[0122] Path complexity: Random selection allows for a sufficient length of path planning, effectively testing the algorithm's path planning capabilities.

[0123] 3. System Operation Flow

[0124] Initialization: The system loads the pre-trained ColorDynamic strategy. The task logic layer generates a double-ended queue Q based on randomly generated orders and calculates the safe offset for the shelf target point.

[0125] Loop control:

[0126] Every 0.1 seconds, the system reads the noisy radar and pose data.

[0127] The logic layer calculates the current target distance. If it is less than the threshold of 0.3m and the queue is not empty, the AGS mechanism is triggered: the old target is popped and the relative pose of the new target is instantly injected into the policy input layer.

[0128] The strategy network outputs speed commands to drive the robot's movement.

[0129] Termination conditions: Reaching the finish line is considered a success; colliding with an obstacle or exceeding 2000 steps is considered a failure.

[0130] 4. Evaluation Indicators and Results

[0131] This embodiment uses the following four indicators for evaluation:

[0132] Task success rate: Number of rounds to successfully complete the task / Total number of rounds.

[0133] Collision rate: Number of turns lost due to collisions with dynamic obstacles / Total number of turns.

[0134] Timeout rate: Number of rounds not completed due to exceeding 2000 steps / Total number of rounds.

[0135] Average number of successful steps: The average number of steps in a successful round, reflecting efficiency.

[0136] Experimental data statistics:

[0137] For three warehouse layouts, task sizes K∈{1, 2, 3, 4, 5} (randomly selecting 1-5 shelves) were set. With a fixed number of 6 obstacles but randomized positions, velocities, and acceleration heights, each configuration ran independently for 100 episodes. The statistical results are shown in Table 2.

[0138] Table 2 Task Performance Comparison Table

[0139]

[0140] Note: Success rate, collision rate, and timeout rate are calculated using Wilson 95% confidence intervals; average number of successful rounds is calculated as mean ± sample standard deviation (SD). All failure types are mutually exclusive (collision, timeout), and the total failure rate = collision rate + timeout rate.

[0141] Considering overall task completion capability, safety performance, planning performance, and execution efficiency, when the task size is 3, under the three layouts, while improving operational efficiency, the task completion capability remains above 84%, collision performance is controlled within 10%, and the timeout rate is controlled within 10%. Under both distributed and centralized simple layouts, when the task size is 3, while improving operational efficiency, the task completion capability can remain above 90%, collision performance is controlled within 10%, and the timeout rate is controlled within 2%.

Claims

1. A dynamic warehouse navigation method that decouples task logic and navigation strategy for single-wave picking, characterized in that: A high-fidelity, highly random, and noisy simulation environment is used as the test framework; A deep reinforcement learning single-target navigation network, ColorDynamic, pre-trained with an independent underlying motion control layer, is used for underlying motion control to realize a scheduling strategy between a single target and dynamic obstacles. An independent upper-level task management layer is used to form an order task queue; During robot navigation, the robot's motion behavior is regulated by the underlying motion control layer, while the tasks are executed continuously based on the multi-objective sequence of the upper task management layer, thereby achieving seamless continuous navigation across multiple objectives.

2. The dynamic warehouse navigation method for decoupling task logic and navigation strategy for single-wave picking as described in claim 1, characterized in that, It includes the following specific steps: Step S1. Construct a high-fidelity dynamic warehouse simulation environment; Step S2. Load the pre-trained underlying navigation policy model; Step S3. Initialize the independent task logic management layer; Step S4. Perform closed-loop control and seamless switching to adaptive target.

3. The dynamic warehouse navigation method for decoupling task logic and navigation strategy for single-wave picking as described in claim 2, characterized in that: The specific method for constructing a high-fidelity dynamic warehouse simulation environment in S1 is as follows: S1-1. Configure the geometric parameters and gait parameters of the warehouse map; S1-2. Initialize a preset number of dynamic obstacle sets in the environment and execute a multi-dimensional randomized perturbation strategy for the warehouse scenario.

4. The dynamic warehouse navigation method for decoupling task logic and navigation strategy for single-wave picking as described in claim 3, characterized in that: The multidimensional randomization perturbation strategy includes: Initial position randomization: The initial positions of dynamic obstacles are uniformly and randomly distributed within the effective area of ​​the map; Initial velocity randomization: the initial velocity vector of the obstacle (v) x , v y Random sampling is performed within a selected range; Randomization of motion behavior: During the motion, the obstacle performs a velocity update once in a selected gait, and the velocity change Δv is randomly selected within a selected range; Geometric size randomization: The physical size of obstacles changes randomly between rounds; Diverse observation noise injection: Superimpose random noise with a uniform distribution onto the robot sensor observation data.

5. A dynamic warehouse navigation method for decoupling task logic and navigation strategy for single-wave picking as described in claim 4, characterized in that: The random noise includes: position measurement noise, direction measurement noise, linear velocity noise, angular velocity noise, and radar measurement noise.

6. A dynamic warehouse navigation method for decoupling task logic and navigation strategy for single-wave picking as described in claim 2, characterized in that: In S2, the specific method for loading the pre-trained underlying navigation policy model is as follows: S2-1. Load an end-to-end navigation policy network ColorDynamic, which is pre-trained based on deep reinforcement learning, and encapsulate it as a black-box execution unit; S2-2. Configure the input vector of the underlying navigation strategy model; The input vector for configuring the underlying navigation strategy model is a noisy sequence of robot states over a specified period of time in the past. The state sequence refers to the set of the robot's relative geometric relationships, the robot's own motion state, and the results of environmental perception at each moment.

7. A dynamic warehouse navigation method for decoupling task logic and navigation strategy for single-wave picking as described in claim 6, characterized in that: The underlying navigation strategy model outputs the selection of discrete actions or the parameters of continuous actions based on the input vector; The underlying motion control layer, based on a dynamic model, transforms the selection of discrete actions or the parameters of continuous actions into linear and angular velocity control signals, thereby regulating the robot's motion behavior.

8. A dynamic warehouse navigation method for decoupling task logic and navigation strategy for single-wave picking as described in claim 2, characterized in that: In S3, the specific method for adding an independent task logic management layer is as follows: construct a task logic management module that is independent of the underlying strategy and is responsible for the task sequence management throughout the entire lifecycle; Among them, including Task queue construction: Based on a single wave of picking orders, randomly select K shelf locations from N map layouts as task targets, and generate a double-ended queue Q containing multiple picking point coordinates and destination coordinates to simulate the order-driven picking path in an actual warehouse.

9. A dynamic warehouse navigation method for decoupling task logic and navigation strategy for single-wave picking as described in claim 2, characterized in that: In S4, the specific method for seamlessly switching between closed-loop control and adaptive objectives is as follows: within each control cycle, the underlying strategy reasoning and the upper-level logic monitoring are executed synchronously. Among them, including State awareness and reasoning: Acquire current noisy observation data and calculate the current target G. curent The relative pose is input into the underlying policy model to obtain motion commands and execute them; Arrival determination: Real-time calculation of the robot's current position and G curent Euclidean distance d t If d t If the value is less than the preset threshold ε, the current subtask is considered complete. Adaptive State Injection: If the current subtask is completed and the number of remaining targets in queue Q is greater than 1, perform a dequeue operation to remove G. curent The new first element of the queue is read as the next target, and the target relative pose parameters in the input vector are instantly updated to the parameters corresponding to the next target. At this time, the underlying strategy, based on the updated target information, directly outputs a smooth motion command pointing to the next target in the next control cycle, realizing the continuous transition between subtasks. Termination judgment: If the queue is cleared and the destination is reached, the task is considered successful; if a collision occurs or the number of steps exceeds the maximum threshold, the task is considered to have failed.

10. A dynamic warehouse navigation system that decouples task logic and navigation strategy for single-wave picking, characterized in that: include The high-fidelity environment construction module executes multi-dimensional randomized perturbation strategies and diverse noise injections for warehousing scenarios. The strategy execution module carries the pre-trained black-box navigation model; The logic control module executes queue management and adaptive target switching logic seamlessly. The data acquisition and analysis module calculates task success rate, collision rate, timeout rate, and average number of successful steps.

Images