Multi-lunar vehicle cooperative path planning and task allocation method for lunar surface

Abstract

The invention discloses a lunar surface-oriented multi-lunar vehicle cooperative path planning and task allocation method, and belongs to the technical field of deep space exploration. The multi-lunar vehicle cooperative path planning and task allocation method solves the problems that a multi-lunar vehicle cooperative path planning and task allocation method is low in environment sensing precision, limited in sensing range, not fused in terrain roughness, resource distribution and multi-machine cooperation, and poor in dynamic environment adaptability. According to the method, the FCN semantic segmentation model, the LSTM model and the CNN model are combined to realize multi-modal environment perception, and constraint conditions are fused in perception data processing, so that the robustness and security of the system in a lunar surface environment with limited perception and incomplete information are improved. And task allocation in a multi-lunar vehicle cooperation process is carried out based on a dynamic environment thermodynamic diagram fusing terrain roughness and resource distribution. And finally, through a staged training strategy, the adaptability and the security of the deep reinforcement learning model in an unknown dynamic environment are improved. The method can be applied to multi-lunar vehicle cooperative path planning and task allocation.

Description

Technical Field

[0001] This invention belongs to the field of deep space exploration technology, specifically relating to a method for collaborative path planning and task allocation for multiple lunar rovers on the lunar surface. Background Technology

[0002] Some highlands in the lunar south pole region enjoy more than 100 consecutive days of sunlight, and water ice may exist in nearby permanently shadowed craters. These conditions are conducive to long-term human and machine operations, making it a strategic location for deep space exploration and the construction of research stations. However, the extreme environment and complex terrain of the lunar surface pose severe challenges to lunar exploration technology.

[0003] In complex dynamic scenarios, centralized scheduling systems are prone to path deadlocks due to communication delays. The lunar surface lacks globally covered communication network infrastructure, and beyond-line-of-sight communication relies on relay satellites, which suffer from significant delays and limited bandwidth. While in-line-of-sight direct communication has low latency, its effective communication distance is short and it is easily disrupted by the rugged terrain of the lunar surface. This uncertainty, limited bandwidth, and potential high latency in communication are key bottlenecks restricting the efficient collaboration of multi-agent systems, especially the reliability of centralized scheduling schemes.

[0004] Furthermore, in traditional lunar rover path planning methods, the rover's actions largely rely on pre-set paths and predefined static maps. However, the lunar rover's operating environment is not entirely known, necessitating a degree of adaptability to cope with environmental changes. Additionally, single sensors are prone to data drift or false detections in low-light, highly reflective lunar environments, leading to insufficient accuracy in identifying icy lunar regolith and inaccurate and untimely perception of surrounding obstacles and dynamic changes. The extreme lunar environment also significantly constrains sensor performance: low temperatures, high radiation, low illumination, or high reflectivity shorten the effective detection range of infrared cameras, limit their viewing angle, and make them susceptible to noise interference; lidar experiences a significant decrease in effective detection range and point cloud density in rugged terrain and low-reflectivity areas, with severe attenuation of edge accuracy; dust accumulation further degrades the performance of optical sensors. These inherent physical constraints severely limit the lunar rover's environmental perception range and accuracy. Furthermore, traditional path planning algorithms, such as A* and RRT, while capable of generating globally optimal paths in static environments, fail to consider terrain ruggedness, resource distribution, and multi-rovers collaborative requirements, leading to excessive system energy consumption or mission conflicts among multiple lunar rovers. Although classic deep reinforcement learning (DRL) techniques improve the dynamic environmental adaptability of lunar rovers through end-to-end decision-making, their convergence efficiency in sparse reward scenarios with uneven resource distribution still needs improvement.

[0005] In summary, to address the problems of low environmental perception accuracy, limited perception range, lack of integration of terrain ruggedness, resource distribution, and multi-robot collaboration in multi-robot collaborative path planning and task allocation methods for lunar surfaces, a new method for multi-robot collaborative path planning and task allocation on the lunar surface is proposed, which is a problem that urgently needs to be solved. Summary of the Invention

[0006] The purpose of this invention is to address the problems of low environmental perception accuracy, limited perception range, lack of integration of terrain ruggedness, resource distribution, and multi-robot collaboration in multi-robot collaborative path planning and task allocation methods for the lunar surface, and to propose a multi-robot collaborative path planning and task allocation method for the lunar surface.

[0007] The technical solution adopted by this invention to solve the above-mentioned technical problems is: a method for collaborative path planning and task allocation for multiple lunar rovers on the lunar surface, the method specifically including the following steps:

[0008] Step 1: Each lunar rover is equipped with an infrared camera, an infrared spectrometer, an IMU, and a lidar.

[0009] Step 2: Generate a dynamic environment heat map for each lunar rover, and each lunar rover transmits its own dynamic environment heat map to the central controller.

[0010] For any lunar rover, the dynamic environment heatmap is generated as follows:

[0011] Step 2: 1. Use the infrared camera on the lunar rover to capture images of the lunar surface. Generate a water content distribution heat map based on the lunar surface images and the scanning results of the infrared spectrometer. Then, generate a distribution map of ice-containing lunar soil resources based on the water content distribution heat map.

[0012] Step 22: Determine the areas containing icy lunar soil based on the distribution map of icy lunar soil resources, then extract the point cloud data of the icy lunar soil areas from the point cloud data obtained by lidar, and calculate the ruggedness index of the icy lunar soil areas based on the extracted point cloud data of the icy lunar soil areas.

[0013] Steps 2 and 3: Downsample the point cloud data collected by the lidar. The remaining points after downsampling correspond to the positions of obstacles in three-dimensional space.

[0014] Step 24: Based on the distribution map of icy lunar regolith resources, identify the lunar regolith areas with an ice content exceeding the threshold, and then generate a dynamic environmental heat map corresponding to the lunar rover based on the location of obstacles, the lunar regolith areas with an ice content exceeding the threshold, and the ruggedness index.

[0015] Step 3: The central controller generates tasks with different priorities based on the received dynamic environmental heat map, and sorts the tasks in descending order of priority to obtain a task list.

[0016] Step 4: Initialization ;

[0017] Step 5: For each lunar rover that has not yet been selected for a mission in the current mission list, calculate its status based on its own condition for the mission in the current mission list. The central controller selects the lunar rover with the highest bid for each mission, thus achieving dynamic mission allocation.

[0018] Step Six, Order Return to step five until all tasks in the task list have been assigned, then proceed to step seven.

[0019] Step 7: Each lunar rover assigned to a task obtains a comprehensive state vector based on its own target acquisition point location and neural network, and then inputs the comprehensive state vector of each lunar rover into a deep reinforcement learning model.

[0020] By using a deep reinforcement learning model to output action commands to each lunar rover performing a task, and then returning to execute step two at the next moment, continuous control of the lunar rover is achieved.

[0021] The beneficial effects of this invention are:

[0022] 1. This invention fully considers the strict constraints that the extreme lunar environment places on sensor performance, including detection distance, viewing angle, and accuracy. By incorporating these constraints into the sensing data processing, environment modeling, and path planning algorithms, the robustness and safety of the system in the lunar environment with limited perception and incomplete information are improved.

[0023] 2. This invention proposes a deep reinforcement learning path planning method based on multi-sensor fusion and semantic segmentation. It utilizes FCN for semantic segmentation of visual images to effectively filter out interference factors. Combined with LSTM temporal modeling of LiDAR point cloud data and the robot's own state information, and CNN spatial feature extraction, multimodal environmental perception is achieved. Furthermore, a deep reinforcement learning model is employed for end-to-end decision-making, improving the real-time performance and obstacle avoidance robustness of the lunar exploration robot in complex lunar environments.

[0024] 3. This invention addresses the problems of poor adaptability and insufficient generalization ability of deep reinforcement learning in dynamic environments by employing a phased training strategy: barrier-free → static obstacle → dynamic obstacle environment, and an adaptive reward function design. Combining the dual-network update mechanism and Q-value decomposition in the deep reinforcement learning model effectively alleviates the Q-value overestimation problem, and accelerates model convergence through priority experience replay, thereby enhancing the system's adaptability and safety in unknown dynamic environments.

[0025] 4. This invention introduces a path planning mechanism that combines dynamic environmental heat maps with three-dimensional terrain modeling. It comprehensively considers terrain ruggedness, lunar soil resource distribution, and dynamic changes in obstacles to achieve a proactive avoidance strategy for high-risk areas and a priority collection strategy for high-value areas. This improves the task completion efficiency and resource utilization of the path planning system and meets the actual needs of high-precision exploration and navigation in complex lunar soil environments.

[0026] 5. This invention proposes a multi-robot collaborative scheduling strategy that combines a bidding mechanism with private blockchain data management. Based on the robot's remaining battery power, location, and task load, the bidding price is dynamically calculated to achieve optimal task allocation and meet the requirements of multi-robot collaboration. At the same time, UWB short-range positioning communication ensures low-latency collaborative operation, and the private blockchain enables closed-loop management of resources such as collection, transportation, and inventory. This significantly improves the coordination and robustness of the multi-robot system in task conflict resolution, resource scheduling, and fault-tolerant operation.

[0027] 6. This invention addresses the communication constraints faced by multi-agent systems on the lunar surface, designing a hierarchical communication and decision-making architecture: normal centralized, weak-communication distributed neighborhood collaboration, and fully offline autonomy. The system does not rely on continuous and stable global communication. Through UWB short-range communication, neighborhood information exchange and relays, a local offline DRL decision model, and private chain data management supporting sharding and caching, it ensures that even under conditions of limited, interrupted, or delayed communication, the multi-robot system can still maintain basic obstacle avoidance safety, task progression capabilities, and data integrity, significantly improving the system's robustness, survivability, and task completion rate in complex lunar environments. Attached Figure Description

[0028] Figure 1 This is an architecture diagram of a multi-lunar rover collaborative path planning and task allocation method for the lunar surface according to the present invention.

[0029] Figure 2 It is a timing diagram of multi-robot collaborative scheduling;

[0030] Figure 3 This is a flowchart of an end-to-end path planning algorithm based on the D3QN PER algorithm;

[0031] Figure 4 This is a schematic diagram of the Q-function network structure;

[0032] Figure 5 This is the training flowchart for the D3QN PER algorithm;

[0033] Figure 6 This is a flowchart of the phased training process. Detailed Implementation

[0034] Specific implementation method one: Combining Figure 1 This embodiment describes a method for collaborative path planning and task allocation for multiple lunar rovers on the lunar surface. The method specifically includes the following steps:

[0035] Step 1: Each lunar rover is equipped with a high-resolution infrared camera, a short-wave infrared spectrometer, an IMU, and a lidar. The infrared camera, infrared spectrometer, and lidar collect data simultaneously, with the IMU's acquisition frequency being higher than that of the infrared camera, infrared spectrometer, and lidar.

[0036] LiDAR is used to collect point cloud data of the surrounding environment; visual cameras acquire lunar surface images and focus on identifying icy areas and obstacles; wheeled odometry combined with IMU is used to estimate the rover's speed, angular velocity, attitude, and heading.

[0037] The parameter settings for each sensor are shown in Table 1:

[0038] Table 1

[0039]

[0040] Step 2: Generate a dynamic environment heat map for each lunar rover, and each lunar rover transmits its own dynamic environment heat map to the central controller.

[0041] For any lunar rover, the dynamic environment heatmap is generated as follows:

[0042] Step 2: In order to capture the infrared reflection characteristics of the icy lunar soil region, the infrared camera on the lunar rover was used to capture lunar surface images. Based on the lunar surface images and the scanning results of the lunar surface by the infrared spectrometer, a water content distribution heat map was generated. Then, based on the water content distribution heat map, a distribution map of icy lunar soil resources was generated.

[0043] The specific process of generating a lunar regolith resource distribution map based on the water content distribution heat map is described as follows: Based on the water content distribution heat map, radiation intensity data is extracted by identifying the characteristic bands of solid water ice at 1.45μm, 1.95μm, and 2.95μm. A mapping function from water content to ice content is established by combining the lunar regolith simulated phase spectrum library JSC-1A calibrated in the laboratory. Based on the mapping function, binarization segmentation is performed to generate an ice-containing resource mask. Finally, the three-dimensional topographic data of the lunar surface is fused to generate a lunar regolith resource distribution map with elevation markers in the lunar-centered and lunar-fixed coordinate system, and high-value collection areas are marked.

[0044] Step 22: Determine the areas containing icy lunar soil based on the distribution map of icy lunar soil resources, then extract the point cloud data of the icy lunar soil areas from the point cloud data obtained by lidar, and calculate the ruggedness index of the icy lunar soil areas based on the extracted point cloud data of the icy lunar soil areas.

[0045] Steps 2 and 3: Downsample the point cloud data collected by the lidar. The remaining points after downsampling correspond to the positions of obstacles in three-dimensional space.

[0046] The specific process of downsampling is as follows: the acquired 3D point cloud data is mapped to the same plane to obtain 2D planar point cloud data; and the obtained 2D planar point cloud data is downsampled. The remaining 2D planar point cloud after downsampling satisfies:

[0047]

[0048] in, and The remaining two-dimensional planar point cloud after downsampling. Indicates the current position of the lunar rover. This represents the Euclidean distance between two points.

[0049] Step 24: Based on the distribution map of icy lunar regolith resources, identify the lunar regolith areas with an ice content exceeding the threshold, and then generate a dynamic environmental heat map corresponding to the lunar rover based on the location of obstacles, the lunar regolith areas with an ice content exceeding the threshold, and the ruggedness index.

[0050] Lunar regolith with high ice content is the priority target area for data collection. This invention defines lunar regolith areas exceeding a threshold of 3 wt% as high ice content areas. Furthermore, it should be noted that the infrared camera, infrared spectrometer, and lidar of this invention can collect data at any given time, then generate a dynamic environmental heat map based on the collected data, thereby completing a path planning operation. Alternatively, the infrared camera, infrared spectrometer, and lidar can collect data at fixed intervals. For dynamic obstacles, if the data collection cycle of the infrared camera, infrared spectrometer, and lidar has not yet arrived, Kalman filtering can be used to predict the trajectory of the dynamic obstacle, updating the dynamic environmental heat map between two data collection intervals, thus guiding path planning.

[0051] Step 3: The central controller generates tasks with different priorities based on the received dynamic environmental heat map, and sorts the tasks in descending order of priority to obtain a task list.

[0052] Step 4: Initialization ;

[0053] Step 5: For each lunar rover that has not yet been selected for a mission in the current mission list, calculate its status based on its own condition for the mission in the current mission list. The central controller selects the lunar rover with the highest bid for each mission, thus achieving dynamic mission allocation.

[0054] Step Six, Order Return to step five until all tasks in the task list have been assigned, then proceed to step seven.

[0055] Step 7: Each lunar rover assigned to a task obtains a comprehensive state vector based on its own target acquisition point location and neural network, and then inputs the comprehensive state vector of each lunar rover into a deep reinforcement learning model.

[0056] The deep reinforcement learning model outputs action commands to each lunar rover performing its tasks. The lunar rover combines the PID control algorithm to calculate the motor control signal, thereby controlling the lunar rover's trajectory at the current moment. This enables the lunar rover to achieve autonomous navigation and dynamic obstacle avoidance, and then returns to execute step two at the next moment, thus achieving continuous control of the lunar rover.

[0057] Specific Implementation Method Two: This implementation method differs from Specific Implementation Method One in that the effective horizontal detection range of the lidar is [missing information]. The sampling frequency is 50 times per second, and the effective detection range in the distance direction is [missing information]. The unit is meters (m).

[0058] Based on point cloud data of the icy lunar soil region collected by lidar, the ruggedness index of the icy lunar soil region is calculated:

[0059]

[0060] in, This is the ruggedness index for regions containing glacial lunar soil. This indicates the number of points in the icy lunar soil region collected by the lidar. The first region containing glacial lunar soil The height of each point This represents the average elevation of all points in the region containing glacial lunar soil.

[0061] like If the value is ≥0.8, the area containing glacial lunar soil is considered high-risk terrain; otherwise, the area containing glacial lunar soil is considered moderately rugged terrain.

[0062] The other steps and parameters are the same as in Specific Implementation Method 1.

[0063] The lidar used in this invention is a 360° lidar, and the height of points outside the detection range of direction and distance dimensions is defaulted to 0 to avoid input noise. Using the method of this embodiment, the ruggedness index of each icy lunar soil region can be calculated separately, and the passability coefficient of pixels within the same icy lunar soil region is the same.

[0064] Specific Implementation Method Three: This implementation method differs from Specific Implementation Method One or Two in that the dynamic environmental heat map is generated in the following way:

[0065] In the distribution map of icy lunar soil resources, areas with an ice content ≥3wt% are marked as high priority, and areas with an ice content <3wt% are marked as low priority. The resource value of pixels in high priority areas is 1, and the resource value of pixels in low priority areas is 0. The passability coefficient of pixels in high-risk terrain areas is 0, and the passability coefficient of pixels in medium-roughness terrain areas is -1. The obstacle coefficient of pixels containing static obstacles is -∞, the obstacle coefficient of pixels containing dynamic obstacles is -10, and the obstacle coefficient of pixels corresponding to non-obstacles is 0. It should be noted that static obstacles and dynamic obstacles are distinguished by transforming the point clouds of two consecutive frames to the same coordinate system. After transforming the point clouds of two consecutive frames to the same coordinate system, the overlapping part of the point cloud is the static obstacle area, and the non-overlapping part of the point cloud is the dynamic obstacle area.

[0066] Calculate the thermal value of each pixel in the dynamic environment thermal map:

[0067] Thermal value = 0.7 × resource value + 0.2 × accessibility coefficient + 0.1 × obstacle coefficient

[0068] The optimal acquisition area is defined as the region containing pixels with a heat value greater than or equal to 0.6 in the dynamic environment heatmap. The area where the pixel is located is designated as the medium-value area. Subsequent data collection tasks will only target the optimal collection area and the medium-value area. After calculating the heat value of each pixel, the entire dynamic environment heat map will be obtained.

[0069] The central controller generates tasks with different priorities based on the received dynamic environmental heatmap, specifically:

[0070] The optimal acquisition area and the medium-value area are divided into various acquisition areas. Each acquisition area corresponds to a task (each task includes two processes: acquisition and transportation). The higher the average heat value of the pixels in the acquisition area, the higher the priority of the task corresponding to the acquisition area.

[0071] Other steps and parameters are the same as in specific implementation method one or two.

[0072] The following is a further explanation of this implementation method:

[0073] 1. Task number setting: Calculate the total area of ​​each continuous area composed of the optimal collection area and the medium value area respectively. Let the total area of ​​any continuous area be A, and let the maximum operating area of ​​the lunar rover in a single month be a.

[0074] like Then, the number N of the regions to be collected obtained by dividing this continuous region is:

[0075]

[0076] in, Indicates rounding up;

[0077] like If so, this continuous area will be directly used as a data collection area;

[0078] After traversing every continuous region in the dynamic environment heatmap corresponding to all lunar rovers, the entire area to be collected is obtained.

[0079] 2. The coordinates of the collection point are: the point with the maximum ice content in each area to be collected.

[0080] The task corresponding to the area with the highest average thermal value of pixels within the region is marked with a priority of 1, the task corresponding to the area with the second highest average thermal value of pixels within the region is marked with a priority of 2, and so on.

[0081] Specific Implementation Method Four: This implementation method differs from Specific Implementation Methods One to Three in that the first... Information for each task includes the data collection point. Lunar soil was collected and transported to the base warehouse.

[0082] The lunar rover calculates the status of the mission in the task list based on its own condition. The bid price for each task is as follows:

[0083]

[0084] in, This indicates the current location of the lunar rover from the data collection point. distance, This indicates the bid price.

[0085] The other steps and parameters are the same as those in one of the specific implementation methods one to three.

[0086] It should be noted that: such as Figure 2 As shown, if multiple lunar rovers bid for the same task simultaneously, priority will be given to assigning the task to the robot with ≥60% remaining battery power. If a single lunar rover cannot be uniquely determined, the rover closest to the data collection point will be selected first. Ultra-wideband (UWB) is used to achieve short-range positioning between robots, with an update frequency of 100Hz, a communication rate of no less than 100 Mbps, and a communication latency of less than 50 ms.

[0087] When the lunar rover maintains a stable UWB connection with the central controller or regional relay node, it periodically reports its status to the central controller, including its position, battery level, and mission status.

[0088] When the connection between the lunar rover and the central controller is unstable or interrupted, but a UWB link can still be established with a neighboring lunar rover (within ≤500m), the rover broadcasts its status, battery level, mission, and local environment map via UWB. A lunar rover with low battery can request a neighboring rover to take over the mission and transmit critical data via UWB. Collaborative operations are recorded on a local private blockchain fragment and synchronized to the central main chain after communication is restored. Lunar rovers with strong communication capabilities or advantageous locations can temporarily act as regional relay nodes to forward critical information, expanding the effective collaboration range.

[0089] When the rover completely loses communication with the central controller and all other rovers, it immediately switches to a pre-loaded local offline DRL decision model. The rover relies solely on its own sensors to continue executing the last valid command for the current mission, and caches all operational and sensing data locally. Once the communication link is restored, rovers in weak communication or offline mode immediately upload their cached local data to the central controller. The central controller integrates the information, updates the global state, and reassigns tasks or adjusts routes as needed. Data cached on the rover's local private blockchain is synchronized to the main blockchain.

[0090] The private blockchain architecture records resource data throughout the entire mission, achieving a closed-loop and tamper-proof information system for each stage of collection, transportation, and storage. Each block records the resource sample coordinates, water content, and collection time; the lunar rover ID, collection path, transportation path, arrival time, time consumed, and energy consumed during the collection mission; and the current inventory, consumption rate, and warning threshold of the research station.

[0091] When the inventory falls below a set threshold, a new round of data collection tasks is automatically generated and assigned to an idle lunar rover. If a lunar rover loses connection with the central controller, it switches to an offline DRL model to continue the current task, synchronizing data once communication is restored. If a lunar rover experiences a motor malfunction (e.g., motor abnormality or path blockage), the central controller identifies the problem through status monitoring, cancels its task assignment, reassigns the task to another lunar rover in normal condition, and instructs the malfunctioning rover to return to the repair point or enter standby mode.

[0092] Specific Implementation Method Five: This implementation method differs from Specific Implementation Methods One to Four in that, in step seven, each lunar rover assigned to a task obtains a comprehensive state vector based on its own target acquisition point location and neural network, specifically as follows:

[0093] For any lunar rover:

[0094] Step 71: First, the FCN model is used as the semantic segmentation model, and the Cityscapes dataset is used to pre-train the FCN model to accurately identify the semantic categories in the images. The lunar surface images captured by the infrared camera on the lunar rover are input into the pre-trained FCN model. The pre-trained FCN model outputs the semantic classification label of each pixel in the lunar surface image to generate a semantic segmentation map.

[0095] Step 72: Set the pixel values ​​of the icy lunar soil region in the semantic segmentation image to 1, and set the pixel values ​​of the non-icy lunar soil region to 0 to obtain a binary mask image. Then, perform a pixel-by-pixel multiplication operation between the binary mask image and the captured lunar surface image to obtain the masked lunar surface image. The masked lunar surface image filters out shadows and background interference, and only retains the effective target region.

[0096] Step 73: Denote the size of the lunar surface image captured by the infrared camera in the current frame after masking as C×C. Stitch the lunar surface image captured by the infrared camera in the current frame after masking and the lunar surface images in the previous 3 frames after masking as a 3D image. The size of the stitched 3D image is C×C×4.

[0097] The stitched 3D image is input into a CNN network with 4 convolutional layers, and the CNN network outputs a 64-dimensional spatial feature vector.

[0098] Step 74: Obtain the lunar rover's own state vector using the IMU. ,in, For the lunar rover in the current The linear velocity at time t, in m / s. For the lunar rover in the current The angular velocity at time t, expressed in rad / s. For the lunar rover in the current The straight-line distance between the location at any given time and the target point, in meters. The angle between the rover's current orientation and the direction of the line connecting the rover and the data collection target point, expressed in rad.

[0099] The rover's own state information guides its movement toward the destination. The rover's own state information at continuous time steps is input into the LSTM network in a time series for real-time feedback control of dynamic path planning.

[0100] Step 75: Establish an LSTM network consisting of L LSTM units. The LSTM network is used to capture the spatial-temporal dependencies of long-term point cloud data, and the value of L can be set according to actual needs. The current frame point cloud collected by the lidar and the point clouds of the previous L-1 consecutive frames are used as inputs to the LSTM network. The LSTM network outputs a 32-dimensional feature vector of the lunar rover's surrounding environment.

[0101] Step 76: Combine the spatial feature vector, the environmental feature vector around the lunar rover, and the coordinates of the collection points assigned to the lunar rover's mission. The vectors are concatenated to form a 98-dimensional integrated state vector.

[0102] The other steps and parameters are the same as those in one of the specific implementation methods one to four.

[0103] like Figure 3 The diagram illustrates the end-to-end path planning framework for a lunar rover based on the D3QN PER algorithm designed in this invention. This method directly maps sensor information to action commands, transforming the rover's environmental perception into an operable navigation strategy. First, sensor data, including the rover's own state, LiDAR point cloud information, and visual images, is acquired. Then, an LSTM network is used to process the rover's state and LiDAR data, while a CNN model analyzes the visual images to identify objects. These data sources are fused using environmental feature fusion technology to create a comprehensive environmental representation. The fused information is used as input to the D3QN PER algorithm, which outputs the optimal path planning decision. This decision is then translated into executable actions by adjusting the rover's angular velocity and speed to achieve smooth navigation.

[0104] Specific Implementation Method Six: This implementation method differs from Specific Implementation Methods One to Five in that the deep reinforcement learning model is trained using a phased, progressive training strategy, specifically as follows:

[0105] Phase 1: Training the deep reinforcement learning model in an accessible environment;

[0106] Phase 2: Training the obstacle avoidance capability of the deep reinforcement learning model in a static obstacle environment;

[0107] The third stage: Training the deep reinforcement learning model to avoid obstacles in a dynamic obstacle environment.

[0108] The other steps and parameters are the same as those in one of the specific implementation methods one to five.

[0109] The phased training process of this invention is as follows: Figure 6 As shown, the deep reinforcement learning model was first trained in an accessible environment. The rover's starting and target points were randomly distributed. The reward function was set as follows: +100 for successfully reaching the target, -0.1 for each step's time penalty, and -0.5 for sudden changes in velocity or angular velocity. Initial training parameters included a learning rate of 0.001 and a batch size of 64. A convergence criterion was set at an average reward greater than 80 over 10 consecutive rounds. Through this first stage of training, the model learned basic navigation skills.

[0110] After the first phase of training converges, the rover enters the second phase of training in a static obstacle environment. A reward system is added to train the rover to avoid obstacles. Positive rewards are given when the rover approaches the data collection target. When the lunar rover is less than 1.5 meters from an obstacle, a penalty of -1 is applied for each step. To ensure smooth rover movement, penalties are applied to changes in velocity and angular velocity. If a collision with an obstacle occurs, a penalty of -50 is applied, and the current turn ends. Priority experience replay is enabled, along with hyperparameters. Setting it to 0.6 controls the priority intensity; it's a hyperparameter. Set the learning rate to 0.4 to balance sampling bias. Reduce the learning rate to 0.0005 to prevent overfitting. Train until the collision rate is below 5% and the average reward is ≥60 points before proceeding to the next training phase.

[0111] In the third phase of training in a dynamic obstacle environment, the reward mechanism is further refined: a reward of +0.2 points is given for each step successfully avoiding a dynamic obstacle (distance > 1 meter), and a penalty of -0.3 points is given for emergency braking. The target network synchronization frequency is adjusted to update once every 500 steps to improve stability. The third phase of training is considered complete when the obstacle avoidance success rate is ≥ 90% and the average reward is ≥ 50 points.

[0112] The training process at each stage is as follows: Figure 5As shown, the trained model is deployed to a physical lunar rover platform for field testing. The test environment simulates various terrain undulations, lighting differences, and obstacle distributions on the lunar surface. Evaluation metrics include: path planning response latency less than 200ms, obstacle avoidance success rate of no less than 85%, and path deviation controlled within 0.5 meters. If the above performance metrics are not met, the model is returned to the first stage for retraining or fine-tuning of the model parameters.

[0113] Specific Implementation Method Seven: This implementation method differs from Specific Implementation Methods One through Six in that the method for calculating the task urgency is as follows:

[0114]

[0115] in, , and All are weighting coefficients. The waiting time for the task is in days. It is the base of the natural logarithm. This indicates the maximum ice content within the area to be sampled.

[0116] The other steps and parameters are the same as those in one of the specific implementation methods one to six.

[0117] In this invention, the weighting coefficient , and The values ​​are 0.5, 0.3, and 0.2.

[0118] Specific Implementation Method Eight: This implementation method differs from one of Specific Implementation Methods One to Seven in that, during the training of the deep reinforcement learning model, the experience replay pool contains the first... Sample The priority calculation method is as follows:

[0119]

[0120] in, Indicates the first in the experience replay pool The priority of each sample It is an arbitrarily small positive number (to avoid zero probability). This indicates taking the absolute value. Indicates the first Temporal difference (TD) error for each sample;

[0121] Timing Differential TD Error for:

[0122]

[0123] in, Indicates the state Select Action The reward Indicates the execution of an action The new state that has been reached Indicates the previous action. This represents the current Q value. Indicates the target Q value. It is a discount factor;

[0124]

[0125] in, This represents the parameters of the current Q-network; to ensure the stability of the algorithm, the parameters of the current Q-network are periodically updated. Synchronize to target Q network parameters This dual-network update mechanism avoids frequent updates to the target Q-value, thereby improving training stability.

[0126] The other steps and parameters are the same as those in any of the specific implementation methods one to seven.

[0127] like Figure 4 The diagram shows a schematic of a Q-function network structure. The Q-value is decomposed into a state-value function. and dominance function Two parts:

[0128]

[0129] Among them, the state value function Used to assess the current state The long-term value; the advantage function Used to measure the state Next, action The relative advantage compared to other actions.

[0130] The combined state vector input to the deep reinforcement learning model passes through two fully connected layer branches, each corresponding to a state-value function. and relative advantage function ,in, This is the current state. For the current action, the action value of each action can be obtained by adding the state value and the advantage value. .

[0131] Specific Implementation Method Nine: This implementation method differs from Specific Implementation Methods One through Eight in that the experience playback pool contains the first... Sampling probability of a sample for:

[0132]

[0133] in, For the first in the experience replay pool The priority of each sample This represents the total number of samples in the experience replay pool. This is a hyperparameter.

[0134] The other steps and parameters are the same as those in one of the specific implementation methods one to eight.

[0135] Hyperparameters Used to control the degree of influence of priorities. When When sampling is completely random; when At this time, sampling is based entirely on priority and can be set according to actual needs. The value of .

[0136] Specific Implementation Method Ten: This implementation method differs from Specific Implementation Methods One to Nine in that the experience playback pool contains the first... Gradient weights of each sample for:

[0137]

[0138] in, It is the size of the experience replay pool. This is a hyperparameter.

[0139] The other steps and parameters are the same as those in any of the specific implementation methods one to nine.

[0140] hyperparameters Used to control the degree of weight adjustment, when When, the weight adjustment is invalid; when At this time, the weight is adjusted entirely based on the priority, and can be set according to actual needs. The value of .

[0141] In summary, this invention first uses a fully convolutional network (FCN) to perform semantic segmentation on the acquired lunar surface infrared images, generating a mask to filter out interference factors such as shadows and backgrounds, while retaining effective targets such as icy lunar soil areas and dynamic or static obstacles. The preprocessed images are then input into a convolutional neural network (CNN) to extract environmental features. Simultaneously, a long short-term memory (LSTM) network is used to process LiDAR point cloud data and the robot's own state information to achieve temporal feature extraction. Next, the multi-dimensional features extracted by CNN and LSTM are fused, and the fused features are input into an improved deep reinforcement learning algorithm. By decomposing the Q-value into a state value function and a dominance function, optimal navigation commands are dynamically generated. The priority of experience playback is adaptively adjusted based on temporal difference (TD) error to accelerate algorithm convergence and avoid overfitting. Finally, a phased progressive training strategy is adopted: Phase 1 trains basic navigation and icy area recognition in an unobstructed environment; Phase 2 trains obstacle avoidance capabilities in a static obstacle environment; and Phase 3 trains obstacle avoidance capabilities in a dynamic obstacle environment. Finally, combined with the real-time task requirements of the research station, a distributed cooperative scheduling algorithm is used to resolve multi-machine path conflicts and manage resources in a closed loop.

[0142] The above examples of the present invention are merely illustrative of the computational model and process of the present invention, and are not intended to limit the implementation of the present invention. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is impossible to exhaustively list all possible implementations here. Any obvious variations or modifications derived from the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A method for collaborative path planning and task allocation for multiple lunar rovers on the lunar surface, characterized in that, The method specifically includes the following steps: Step 1: Each lunar rover is equipped with an infrared camera, an infrared spectrometer, an IMU, and a lidar. Step 2: Generate a dynamic environment heat map for each lunar rover, and each lunar rover transmits its own dynamic environment heat map to the central controller. For any lunar rover, the dynamic environment heatmap is generated as follows: Step 2:

1. Use the infrared camera on the lunar rover to capture images of the lunar surface. Generate a water content distribution heat map based on the lunar surface images and the scanning results of the infrared spectrometer. Then, generate a distribution map of ice-containing lunar soil resources based on the water content distribution heat map. Step 22: Determine the areas containing icy lunar soil based on the distribution map of icy lunar soil resources, then extract the point cloud data of the icy lunar soil areas from the point cloud data obtained by lidar, and calculate the ruggedness index of the icy lunar soil areas based on the extracted point cloud data of the icy lunar soil areas. Steps 2 and 3: Downsample the point cloud data collected by the lidar. The remaining points after downsampling correspond to the positions of obstacles in three-dimensional space. Step 24: Based on the distribution map of icy lunar regolith resources, identify the lunar regolith areas with an ice content exceeding the threshold, and then generate a dynamic environmental heat map corresponding to the lunar rover based on the location of obstacles, the lunar regolith areas with an ice content exceeding the threshold, and the ruggedness index. Step 3: The central controller generates tasks with different priorities based on the received dynamic environmental heat map, and sorts the tasks in descending order of priority to obtain a task list. Step 4: Initialization ; Step 5: For each lunar rover that has not yet been selected for a mission in the current mission list, calculate its status based on its own condition for the mission in the current mission list. The central controller selects the lunar rover with the highest bid for each mission, thus achieving dynamic mission allocation. Step Six, Order Return to step five until all tasks in the task list have been assigned, then proceed to step seven. Step 7: Each lunar rover assigned to a task obtains a comprehensive state vector based on its own target acquisition point location and neural network, and then inputs the comprehensive state vector of each lunar rover into a deep reinforcement learning model. By using a deep reinforcement learning model to output action commands to each lunar rover performing a task, and then returning to execute step two at the next moment, continuous control of the lunar rover is achieved.

2. The method for multi-lunar rover cooperative path planning and task allocation for the lunar surface as described in claim 1, characterized in that, The effective horizontal detection range of the lidar is: The sampling frequency is 50 times per second, and the effective detection range in the distance direction is [missing information]. The unit is meters (m). Based on point cloud data of the icy lunar soil region collected by lidar, the ruggedness index of the icy lunar soil region is calculated: in, This is the ruggedness index for regions containing glacial lunar soil. This indicates the number of points in the icy lunar soil region collected by the lidar. The first region containing glacial lunar soil The height of each point This represents the average elevation of all points in the region containing glacial lunar soil. like If the value is ≥0.8, the area containing glacial lunar soil is considered high-risk terrain; otherwise, the area containing glacial lunar soil is considered moderately rugged terrain.

3. The method for multi-lunar rover cooperative path planning and task allocation for the lunar surface according to claim 2, characterized in that, The dynamic environmental heat map is generated as follows: Calculate the thermal value of each pixel in the dynamic environment thermal map: Thermal value = 0.7 × resource value + 0.2 × accessibility coefficient + 0.1 × obstacle coefficient The optimal acquisition area is defined as the region containing pixels with a heat value greater than or equal to 0.6 in the dynamic environment heatmap. The region containing the pixel is designated as the medium-value region; The central controller generates tasks with different priorities based on the received dynamic environmental heatmap, specifically: The optimal acquisition area and the medium-value area are divided into various acquisition areas. Each acquisition area corresponds to a task. The higher the average heat value of the pixels in the acquisition area, the higher the priority of the task corresponding to the acquisition area.

4. The method for multi-lunar rover cooperative path planning and task allocation for the lunar surface as described in claim 3, characterized in that, The first Information for each task includes the data collection point. Lunar soil was collected and transported to the base warehouse. The lunar rover calculates the status of the mission in the task list based on its own condition. The bid price for each task is as follows: in, This indicates the current location of the lunar rover from the data collection point. distance, This indicates the bid price.

5. A method for collaborative path planning and task allocation for multiple lunar rovers on the lunar surface according to claim 4, characterized in that, In step seven, each lunar rover assigned a task obtains a comprehensive state vector based on its own target acquisition point location and neural network, specifically: For any lunar rover: Step 71: Input the lunar surface images captured by the infrared camera on the lunar rover into the pre-trained FCN model, and output the semantic classification label of each pixel in the lunar surface image through the pre-trained FCN model to generate a semantic segmentation map. Step 72: Set the pixel values ​​of the icy lunar soil regions in the semantic segmentation image to 1, and set the pixel values ​​of the non-icy lunar soil regions to 0 to obtain a binary mask image. Then, perform a pixel-by-pixel multiplication operation between the binary mask image and the captured lunar surface image to obtain the masked lunar surface image. Step 73: Denote the size of the lunar surface image captured by the infrared camera in the current frame after masking as C×C. Stitch the lunar surface image captured by the infrared camera in the current frame after masking and the lunar surface images in the previous 3 frames after masking as a 3D image. The size of the stitched 3D image is C×C×4. The stitched 3D image is input into a CNN network, which outputs a 64-dimensional spatial feature vector. Step 74: Obtain the lunar rover's own state vector using the IMU. ,in, For the lunar rover in the current linear velocity at time t, For the lunar rover in the current angular velocity at time t, For the lunar rover in the current The straight-line distance between the location at any given time and the target point. The angle between the rover's current orientation and the direction of the line connecting the rover and the data collection target point; Step 75: Establish an LSTM network consisting of L LSTM units. The point cloud of the current frame collected by the lidar and the point clouds of the previous L-1 consecutive frames are used as the input of the LSTM network. The LSTM network outputs a 32-dimensional feature vector of the lunar rover's surrounding environment. Step 76: Combine the spatial feature vector, the environmental feature vector around the lunar rover, and the coordinates of the collection points assigned to the lunar rover's mission. The vectors are concatenated to form a 98-dimensional integrated state vector.

6. A method for multi-lunar rover cooperative path planning and task allocation for the lunar surface according to claim 5, characterized in that, The deep reinforcement learning model is trained using a phased, progressive training strategy, specifically: Phase 1: Training the deep reinforcement learning model in an accessible environment; Phase 2: Training the obstacle avoidance capability of the deep reinforcement learning model in a static obstacle environment; The third stage: Training the deep reinforcement learning model to avoid obstacles in a dynamic obstacle environment.

7. A method for collaborative path planning and task allocation for multiple lunar rovers on the lunar surface according to claim 6, characterized in that, The method for calculating the urgency of the task is as follows: in, , and All are weighting coefficients. The waiting time for the task is in days. It is the base of the natural logarithm. This indicates the maximum ice content within the area to be sampled.

8. A method for collaborative path planning and task allocation for multiple lunar rovers on the lunar surface according to claim 7, characterized in that, During the training of the deep reinforcement learning model, the experience replay pool contains the first... Sample The priority calculation method is as follows: in, Indicates the first in the experience replay pool The priority of each sample For any small positive number, This indicates taking the absolute value. Indicates the first Temporal difference (TD) error for each sample; Timing Differential TD Error for: in, Indicates the state Select Action The reward Indicates the execution of an action The new state that has been reached Indicates the previous action. This represents the current Q value. Indicates the target Q value. It is a discount factor; in, This represents the parameters of the current Q-network.

9. A method for collaborative path planning and task allocation for multiple lunar rovers on the lunar surface according to claim 8, characterized in that, The experience replay pool of the first Sampling probability of a sample for: in, For the first in the experience replay pool The priority of each sample This represents the total number of samples in the experience replay pool. This is a hyperparameter.

10. A method for multi-lunar rover cooperative path planning and task allocation for the lunar surface according to claim 9, characterized in that, The experience replay pool of the first Gradient weights of each sample for: in, It is the size of the experience replay pool. This is a hyperparameter.

Images