Robot autonomous target searching method oriented to unknown environment

Abstract

The invention discloses a robot autonomous target searching method for an unknown environment, and belongs to the field of artificial intelligence and robot control. The method comprises the following steps: converting environment data collected by a robot at a current position into point cloud, generating an observed area, detecting a target in the observed area, and outputting the position and confidence of the target; meanwhile, based on the observed area, an unobserved area is iteratively predicted through a pre-training generation model, and the probability of occurrence of a target is counted. And fusing the information of the two to construct a comprehensive belief state, simulating the movement of the robot in a discrete action space by using an Image-POUCT algorithm based on the state, and scoring and selecting an optimal movement position through a dense reward function. The process is continuously iterated until a sufficient number of targets with the confidence reaching the standard are found, and all target positions are output; according to the invention, efficient intelligent search of the robot in a large-range unknown environment can be realized.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and robot control, and in particular to a method for autonomous target search of robots in unknown environments. Background Technology

[0002] In applications such as disaster relief and scientific research, autonomous target search technology for robots is of great value. However, existing methods generally suffer from systemic problems of limited environmental perception and low decision-making and planning efficiency when facing large-scale, complex, and unknown environments, which seriously restricts their application effectiveness in practical tasks.

[0003] In terms of environmental perception, existing technologies have significant limitations. Among mainstream real-time localization and mapping (SLAM) technologies, laser SLAM is limited by sensor detection distance and can only build maps of limited areas; visual SLAM is sensitive to changes in lighting and has difficulty directly obtaining accurate scale information, preventing robots from establishing a complete global environmental understanding. This limitation in perception directly affects the effectiveness of subsequent decision-making and planning.

[0004] At the decision-making and planning level, traditional methods struggle to effectively handle state estimation in unobserved regions. While end-to-end methods achieve integrated perception and decision-making, they heavily rely on large amounts of labeled data and exhibit poor environmental adaptability. Modular methods based on POMDP provide a theoretical framework, but their belief state updates depend entirely on historical observation data, lacking effective reasoning capabilities for completely unobserved regions. This disconnect between the perception and decision-making modules leads to significant blind spots in the robot's exploration process, reducing search efficiency and increasing the risk of missing key targets. Therefore, there is an urgent need for an autonomous target search method for robots in unknown environments that can overcome the limitations of local perception and possess reasoning capabilities for unobserved regions. Summary of the Invention

[0005] The purpose of this invention is to provide a robot autonomous target search method for unknown environments, which enables robots to perform efficient and intelligent searches in a wide range of unknown environments.

[0006] To address the aforementioned technical problems, embodiments of the present invention provide a robot autonomous target search method for unknown environments, comprising the following steps: The unknown environment observation data collected by the robot at the current location is converted into point cloud and aggregated to generate an observed area. If an observed target is detected from the observed area, the location information and confidence level of the observed target are output. Based on the observed area, multiple unobserved areas are generated sequentially from near to far using a pre-trained generative model. By counting the number of times the unobserved target appears in the multiple unobserved areas, the location and probability of the unobserved target are obtained. By integrating the location information and confidence level of observed targets, and the existence location and probability of unobserved targets, a comprehensive belief state reflecting a complete understanding of the preset environment is constructed. Based on the comprehensive belief state, the Imagine-POUCT algorithm is used to simulate all possible movement positions of the robot in the discrete action space. Each movement position is scored using a dense reward function, and the robot's highest-scoring movement position is determined. The Imagine-POUCT algorithm, based on the tree search strategy of the PUCT algorithm, recreates the search tree based on the new comprehensive belief state after each robot movement. The above steps are executed iteratively until a preset number of observed targets are successfully found, and the confidence level of these observed targets is greater than a preset threshold. Then the iteration stops, and the location information of all observed targets is output.

[0007] In some optional embodiments, converting the unknown environment observation data into point clouds and aggregating them to generate observed areas includes the following steps: The observation data of the unknown environment is extracted into point clouds of different colors using the SLAM method, and the different colors correspond to different objects; The point cloud is clustered based on the data point density using the DBSCAN algorithm. By enhancing the spatial continuity of point clouds of the same color, point clouds of the same color are made into independent clusters, generating an observed area scene. The observed area scene is wrapped by a bounding box around each independent cluster, forming a spatial location and contour description of different objects.

[0008] In some optional embodiments, the step of generating multiple unobserved regions from near to far using a pre-trained generative model based on the observed region, and obtaining the location and probability of the unobserved target by counting the occurrence frequency of the unobserved target in the multiple unobserved regions, includes the following steps: The observed region is converted into a discrete three-dimensional voxel grid, and each voxel unit in the grid is assigned a semantic label. The unobserved region is marked as an unknown state. Based on a three-dimensional voxel mesh, a pre-trained diffusion model is used, along with a redrawing technique, to iteratively denoise the noise in the unobserved region using voxels of the observed region as a condition, and to generate multiple unobserved regions from near to far in an iterative manner. Based on multiple unobserved regions, the locations of unobserved targets are extracted using the DBSCAN clustering algorithm, and the probability of target presence at each location is calculated using the Monte Carlo statistical method.

[0009] In some alternative embodiments, the discrete motion space defines 12 candidate motions, each of which is a two-dimensional displacement vector on a horizontal plane. These vectors share a fixed step size and are uniformly distributed at 30-degree intervals within a 360-degree range.

[0010] Embodiments of the present invention also provide an autonomous target search system for robots in unknown environments, comprising: Sensing module: Converts the unknown environment observation data collected by the robot at the current position into point cloud and aggregates it to generate the observed area. If the observed target is detected in the observed area, the position information and confidence level of the observed target are output. Imagination Module: Based on the observed area, a pre-trained generative model iteratively generates multiple unobserved areas from near to far. By counting the number of times the unobserved target appears in the multiple unobserved areas, the location and probability of the unobserved target are obtained. Decision Module: By integrating the location information and confidence level of observed targets, and the existence location and probability of unobserved targets, a comprehensive belief state reflecting a complete understanding of the preset environment is constructed. Based on the comprehensive belief state, the Imagine-POUCT algorithm is used to simulate all possible movement positions of the robot in the discrete action space. Each movement position is scored using a dense reward function, and the robot's highest-scoring movement position is determined. The Imagine-POUCT algorithm, based on the tree search strategy of the PUCT algorithm, recreates the search tree based on the new comprehensive belief state after each robot movement. Perception-Imagination-Decision Loop Module: Iterate through the above steps until a preset number of observed targets are successfully found, and the confidence level of these observed targets is greater than a preset threshold. Then stop iterating and output the location information of all observed targets.

[0011] Embodiments of the present invention also provide a computer device, including: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the above-described robot autonomous target search method for unknown environments.

[0012] Embodiments of the present invention also provide a computer-readable storage medium storing a computer program, which, when run by a processor, is capable of executing the above-described robot autonomous target search method for unknown environments.

[0013] The autonomous target search method for robots in unknown environments provided by this invention has at least the following beneficial effects: This invention achieves efficient intelligent search for robots in unknown environments by constructing a closed-loop system of perception, imagination, and decision-making: the perception module converts observation data of the unknown environment into structured point cloud information, establishing an accurate understanding of the observed area; the imagination module innovatively uses a diffusion model to iteratively predict unobserved areas from near to far, generating a probability distribution of unobserved targets, breaking through the limitation of traditional methods that rely solely on historical observation data; the decision-making module integrates observed information and prediction results to construct a comprehensive belief state and uses an improved Imagine-POUCT algorithm for dynamic planning, enabling the robot to achieve an optimal balance between exploration and utilization. This system significantly improves search efficiency and success rate in complex environments through continuous iterative optimization of the search strategy, while ensuring the reliability of search results through a confidence threshold mechanism. Attached Figure Description

[0014] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this invention, illustrate exemplary embodiments of the invention and are used to explain the invention, but do not constitute an undue limitation of the invention. In the drawings:

[0015] Figure 1 This is a flowchart of a robot autonomous target search method for unknown environments provided according to an embodiment of the present invention; Figure 2 This is a schematic diagram of a close-up scene generation process in an unobserved area according to an embodiment of the present invention. Figure 2 (a) is the original image of the initial close-up scene. Figure 2 (b) is a core region map after 4-neighborhood cropping. Figure 2 (c) is a complete mid-to-close range scene image after global fusion; Figure 3 This is a schematic diagram of a long-distance scene generation process in an unobserved area according to an embodiment of the present invention. Figure 3 (a) is the initial mid-to-long distance scene. Figure 3 (b) is a core region map after 9-neighborhood cropping. Figure 3 (c) is a complete mid-to-long-range scene map after global fusion; Figure 4 This is a schematic diagram of the DBSCAN clustering algorithm and Monte Carlo statistical method according to an embodiment of the present invention. Figure 4 (a) is the original scene image before clustering. Figure 4 (b) is a cluster distribution diagram after DBSCAN clustering. Figure 4 (c) is a structured scene diagram after Monte Carlo statistics; Figure 5 This is a schematic diagram of the intelligent agent's environmental task execution process, which involves the coordinated operation of three modules: perception, imagination, and decision-making, according to an embodiment of the present invention. Detailed Implementation

[0016] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0017] One embodiment of the present invention relates to a robot autonomous target search method for unknown environments. The implementation details of the robot autonomous target search method for unknown environments in this embodiment are described in detail below. The following implementation details are provided for ease of understanding and are not necessary for implementing this solution.

[0018] The specific process of the robot autonomous target search method for unknown environments in this embodiment can be described as follows: Figure 1 As shown, it includes: Step 101: Convert the unknown environment observation data collected by the robot at the current location into point cloud and aggregate it to generate an observed area. If an observed target is detected from the observed area, output the location information and confidence level of the observed target. The observation data of the unknown environment is extracted into point clouds of different colors using the SLAM method, and the different colors correspond to different objects; The point cloud is clustered based on the data point density using the DBSCAN algorithm. By enhancing the spatial continuity of point clouds of the same color, point clouds of the same color are made into independent clusters, generating an observed area scene. The observed area scene is wrapped by a bounding box around each independent cluster, forming a spatial location and contour description of different objects.

[0019] Step 102: Based on the observed area, multiple unobserved areas are generated sequentially from near to far using a pre-trained generative model. The location and probability of the unobserved target are obtained by counting the number of times the unobserved target appears in the multiple unobserved areas. The observed region is converted into a discrete three-dimensional voxel grid, and each voxel unit in the grid is assigned a semantic label. The unobserved region is marked as an unknown state. Based on a three-dimensional voxel mesh, a pre-trained diffusion model is used, along with a redrawing technique, to iteratively denoise the noise in the unobserved region using voxels of the observed region as a condition, and to generate multiple unobserved regions from near to far in an iterative manner. Based on multiple unobserved regions, the locations of unobserved targets are extracted using the DBSCAN clustering algorithm, and the probability of target presence at each location is calculated using the Monte Carlo statistical method.

[0020] The process of generating close-up scenes in unobserved areas is as follows: Figure 2 As shown in the figure, the initial mid-to-close range scene (including basic structure and color distribution) is first obtained as shown in Figure (a). Then, Figure (b) is obtained through a 4-neighborhood cropping operation. The core area is located by crosshairs, key structures are preserved and redundant edge information is cropped. Finally, Figure (c) is obtained through global fusion processing. The cropped core area is integrated with the surrounding environmental information to form a complete mid-to-close range scene with rich details.

[0021] The process of generating distant scenes in unobserved areas is as follows: Figure 3 As shown in the figure, the initial mid-to-long-range scene shown in Figure (a) is obtained first, including the corresponding structure and color distribution; then, after 9-neighborhood cropping, the core area is located by dashed lines, the key structure of the scene is preserved and redundant information is simplified to obtain Figure (b); finally, through global fusion, the cropped core area is integrated with the surrounding environment information to enrich the level of detail and form the complete mid-to-long-range scene shown in Figure (c).

[0022] DBSCAN clustering algorithm and Monte Carlo statistical methods, such as Figure 4 As shown in the figure, the scene data in Figure (a) is used as a basis to classify the scattered elements into clusters through DBSCAN clustering, resulting in Figure (b) which focuses on the core area; then, the Monte Carlo statistical method is used to analyze and process the clustered areas, finally forming the structured statistical scene result shown in Figure (c).

[0023] Step 103: By integrating the location information and confidence level of observed targets, and the existence location and probability of unobserved targets, a comprehensive belief state reflecting the complete cognition of the preset environment is constructed. Based on the comprehensive belief state, the Imagine-POUCT algorithm is used to simulate all possible movement positions of the robot in the discrete action space. Each movement position is scored using a dense reward function, and the robot's highest-scoring movement position is determined. The Imagine-POUCT algorithm, based on the tree search strategy of the PUCT algorithm, recreates the search tree based on the new comprehensive belief state after each robot movement. The discrete action space defines 12 candidate actions, each of which is a two-dimensional displacement vector on a horizontal plane. These vectors share a fixed step size and are uniformly distributed in 30-degree intervals within a 360-degree range.

[0024] The Imagine-POUCT algorithm is a dynamic loop system integrating multi-source information processing and intelligent decision-making. The specific process is as follows: First, the root belief state, initial observations, and point cloud data are initialized. Then, within a preset maximum time step, closed-loop operations of environmental perception and action decision-making are executed sequentially in chronological order. In each time step, the sensor first collects the direct perceived belief of the current environment, then combines it with the belief state of the previous moment, the environment model, the transition model, and historical observation results. Through a belief update mechanism, a belief state derived from historical information is obtained. Simultaneously, the imagination module is invoked to generate a virtual belief state using the previous moment's observation as input. Then, the sensor-perceived belief, the historically updated belief, and the virtual imagined belief are merged to form the comprehensive belief state of the current moment, which is used as the root node of the search subtree. Finally, within the specified planning time, POUCT is used... The algorithm iteratively calculates the search subtree, continuously optimizes the value estimate of each possible action, and selects the action with the highest value to execute after the planning time ends. This allows the algorithm to obtain new environmental observation results and agent state, providing input for the next time step cycle. Ultimately, the algorithm realizes the complete process of continuous optimization of decision-making by the agent in a dynamic and partially observable environment through multi-source information fusion and efficient planning.

[0025] Step 104: Iterate through the above steps until a preset number of observed targets are successfully found, and the confidence level of these observed targets is greater than a preset threshold. Then stop iterating and output the location information of all observed targets.

[0026] A schematic diagram of the intelligent agent's environmental task execution process, which involves the coordinated operation of the three modules of perception, imagination, and decision-making, is shown below. Figure 5 As shown in the figure, the agent first obtains RGB, LiDAR, and semantic segmentation information from the unknown environment, which is then processed into scene data by the semantic voxel perception module. Next, the imagination module processes the scene data through a two-layer diffusion model. After that, the decision module first processes the scene data with clustering-Monte Carlo estimation, and then obtains the optimal action through POMDP modeling and POUCT calculation. Finally, the agent executes the action in the simulation environment.

[0027] To objectively measure the effectiveness of the method of this invention, the following three types of baseline methods are designed for comparison: Fully Observable Markov Decision Process (MDP): This method assumes that the robot has complete information about the global environment before performing the task, including the precise locations of all obstacles and targets, thus eliminating the partial observability problem. Its decision-making process is based on a deterministic transfer function and a known reward function, requiring only SLAM odometry for self-localization and the use of the UCT algorithm to solve for the optimal action. This baseline serves as a performance upper limit reference under ideal conditions.

[0028] Partially Observable Markov Decision Process (POMDP): This method is a classic framework for handling partially observable environments, relying solely on real-time sensor observation data. It describes environmental uncertainty by updating belief states (i.e., the probability distribution of states) using Bayesian rules. This baseline does not integrate the "imagination module" of this invention, making it unable to predict unobserved areas, resulting in incomplete belief estimation. Its action selection depends entirely on current observations and historical beliefs. This baseline is used to verify the necessity of introducing the "imagination module."

[0029] Human test subjects: To reflect the level of human intuitive decision-making, 10 participants with computer operation experience (average age 25±3 years, no professional knowledge in the field of robotics) were randomly recruited. They controlled the robot from a first-person perspective through a 3D interactive interface, and could choose a movement action from 12 directions consistent with the algorithm for each step. The interface only displayed the sensor observation range (a fan-shaped area with a radius of 20 meters) in real time. Participants could not obtain a global map, but could only make decisions by constructing a mental map based on historical observations, thus simulating the human "perception-decision-making" process.

[0030] This embodiment sets up Example 1: a T-junction area with a total area of ​​approximately 40,000 m², containing multiple T-junctions, as well as static obstacles such as buildings, vegetation, and streetlights. Eight vehicles to be searched are randomly distributed in the environment as targets, all located near buildings.

[0031] This embodiment sets up two typical scenarios in Example 2: a straight road area. This area is a single straight road with low buildings and sparse vegetation on both sides, with a total area of ​​approximately 30,000 m². Five vehicles to be searched are distributed near the buildings.

[0032] To quantitatively evaluate the performance of this method in instances, the following three types of metrics are introduced: Coverage measures the completeness of the target search and reflects the algorithm's ability to detect targets globally. Its calculation formula is:

[0033] ; in The target number of successful detections. The total number of targets is preset in the environment. In large-scale environments, targets are discretely distributed, and CR (Target Comparison) can directly reflect whether the algorithm has missed key targets, making it a core indicator for evaluating task completion.

[0034] Efficiency is used to quantify the target discovery capability per unit action, revealing the resource utilization efficiency of the algorithm in action planning. Its calculation formula is:

[0035] ; in This refers to the total number of actions performed by the robot. In practical applications, robots have limited energy and time, making efficient strategies (such as fewer actions and more discovery) more valuable in engineering, as they can avoid the "blind wandering" problem of traditional methods.

[0036] The effective exploration rate characterizes the effectiveness of exploration actions, distinguishing between "goal-oriented exploration" and "meaningless traversal." Its calculation formula is:

[0037] ; in The number of targets discovered through exploring new areas. To explore new regions, the number of actions is considered. Traditional POMDP may repeatedly visit known regions, while EER focuses on the algorithm's ability to model unobserved regions by eliminating invalid actions. This is particularly suitable for evaluating the scientific validity of exploration guided by "imagination modules".

[0038] The comparative experiments of Example 1 are shown in Table 1, and the comparative experiments of Example 2 are shown in Table 2: Table 1 Comparative experiments of Example 1; This invention demonstrates significant improvements across all metrics, proving the effectiveness of the imagination module in exploring complex intersection environments. Compared to human test subjects (CR=0.68, E=0.28, EER=0.36), this method outperforms in exploration efficiency (E) and target discovery relevance (EER). The MDP baseline (CR=0.73, E=0.24, EER=0.60), representing the theoretical upper limit, best matches the performance of this method.

[0039] Table 2 Comparative experiments of Example 2; The imagination module of this invention can more accurately predict the distribution of targets, thereby guiding the robot to explore efficiently and avoiding the blindness of POMDP and the instability of human decision-making.

[0040] This invention demonstrates significant performance in target search. Experimental data shows that in simple scenarios such as straight roads, the target coverage rate reaches 0.84, and in complex environments such as T-junctions, it maintains a coverage rate of 0.61, representing a 96.8% improvement over the traditional POMDP method and a clear advantage over human testers. This improvement is mainly attributed to the ability of the imagination module to generate unobserved areas through a diffusion model, enabling the system to predict the target distribution probability based on environmental features, thereby guiding more reasonable search path planning.

[0041] In terms of search efficiency, this invention achieves an efficiency index of 0.32 in simple scenarios, a 113% improvement over the 0.15 of human testers. This is attributed to the dynamic balance mechanism established by the dense reward function between exploring unknown regions and utilizing known information. Of particular note is the system's effective exploration rate of 0.47-0.67, meaning that nearly half of the exploration actions successfully discover new targets, far exceeding the 0.40-0.43 of the traditional POMDP method. This verifies that the unobserved target probability distribution generated by the imagination module through DBSCAN clustering and Monte Carlo statistics can effectively guide the Imagine-POUCT algorithm to prioritize exploration of high-probability regions, thereby significantly improving the targeting and efficiency of the search.

[0042] The embodiments of the present invention also provide an autonomous target search system for robots in unknown environments. The implementation details of the autonomous target search system for robots in unknown environments in this embodiment are described in detail below. The following content is only for the convenience of understanding the implementation details and is not necessary for implementing this solution.

[0043] Specifically, the perception module converts the unknown environment observation data collected by the robot at its current location into point clouds and aggregates them to generate an observed area. If an observed target is detected in the observed area, the module outputs the location information and confidence level of the observed target. Imagination Module: Based on the observed area, a pre-trained generative model iteratively generates multiple unobserved areas from near to far. By counting the number of times the unobserved target appears in the multiple unobserved areas, the location and probability of the unobserved target are obtained. Decision Module: By integrating the location information and confidence level of observed targets, and the existence location and probability of unobserved targets, a comprehensive belief state reflecting a complete understanding of the preset environment is constructed. Based on the comprehensive belief state, the Imagine-POUCT algorithm is used to simulate all possible movement positions of the robot in the discrete action space. Each movement position is scored using a dense reward function, and the robot's highest-scoring movement position is determined. The Imagine-POUCT algorithm, based on the tree search strategy of the PUCT algorithm, recreates the search tree based on the new comprehensive belief state after each robot movement. Perception-Imagination-Decision Loop Module: Iterate through the above steps until a preset number of observed targets are successfully found, and the confidence level of these observed targets is greater than a preset threshold. Then stop iterating and output the location information of all observed targets.

[0044] The steps of the various methods described above are only for clarity. In practice, they can be combined into one step or some steps can be split into multiple steps. As long as they include the same logical relationship, they are all within the protection scope of this invention. Adding insignificant modifications or introducing insignificant designs to the algorithm or process, without changing the core design of the algorithm and process, are also within the protection scope of this invention.

[0045] Another embodiment of the present invention relates to a computer-readable storage medium storing a computer program. When executed by a processor, the computer program implements the method embodiments described above.

[0046] That is, those skilled in the art will understand that all or part of the steps in the methods of the above embodiments can be implemented by a program instructing related hardware. This program is stored in a storage medium and includes several instructions to cause a device (which may be a microcontroller, chip, etc.) or processor to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0047] Those skilled in the art will understand that the above embodiments are specific embodiments for implementing the present invention, and in practical applications, various changes can be made to them in form and detail without departing from the spirit and scope of the present invention.

Claims

1. A method for autonomous target search of robots in unknown environments, characterized in that, The method includes: The unknown environment observation data collected by the robot at the current location is converted into point cloud and aggregated to generate an observed area. If an observed target is detected from the observed area, the location information and confidence level of the observed target are output. Based on the observed area, multiple unobserved areas are generated sequentially from near to far using a pre-trained generative model. By counting the number of times the unobserved target appears in the multiple unobserved areas, the location and probability of the unobserved target are obtained. By integrating the location information and confidence level of observed targets, and the existence location and probability of unobserved targets, a comprehensive belief state reflecting a complete understanding of the preset environment is constructed. Based on the comprehensive belief state, the Imagine-POUCT algorithm is used to simulate all possible movement positions of the robot in the discrete action space. Each movement position is scored using a dense reward function, and the robot's highest-scoring movement position is determined. The Imagine-POUCT algorithm, based on the tree search strategy of the PUCT algorithm, recreates the search tree based on the new comprehensive belief state after each robot movement. The above steps are executed iteratively until a preset number of observed targets are successfully found, and the confidence level of these observed targets is greater than a preset threshold. Then the iteration stops, and the location information of all observed targets is output.

2. The robot autonomous target search method for unknown environments as described in claim 1, characterized in that, The process of converting the unknown environment observation data into point clouds and aggregating them to generate the observed region includes the following steps: The observation data of the unknown environment is extracted into point clouds of different colors using the SLAM method, and the different colors correspond to different objects; The point cloud is clustered based on the data point density using the DBSCAN algorithm. By enhancing the spatial continuity of point clouds of the same color, point clouds of the same color are made into independent clusters, generating an observed area scene. The observed area scene is wrapped by a bounding box around each independent cluster, forming a spatial location and contour description of different objects.

3. The robot autonomous target search method for unknown environments as described in claim 1, characterized in that, Based on the observed region, a pre-trained generative model iteratively generates multiple unobserved regions from near to far. By counting the occurrence frequency of unobserved targets in these unobserved regions, the location and probability of the unobserved targets are obtained. This includes the following steps: The observed region is converted into a discrete three-dimensional voxel grid, and each voxel unit in the grid is assigned a semantic label. The unobserved region is marked as an unknown state. Based on a three-dimensional voxel mesh, a pre-trained diffusion model is used, along with a redrawing technique, to iteratively denoise the noise in the unobserved region using voxels of the observed region as a condition, and to generate multiple unobserved regions from near to far in an iterative manner. Based on multiple unobserved regions, the locations of unobserved targets are extracted using the DBSCAN clustering algorithm, and the probability of target presence at each location is calculated using the Monte Carlo statistical method.

4. The robot autonomous target search method for unknown environments as described in claim 1, characterized in that, The discrete action space defines 12 candidate actions, each of which is a two-dimensional displacement vector on a horizontal plane. These vectors share a fixed movement step size and are evenly distributed in 30-degree intervals within a 360-degree range.

5. A robot autonomous target search system for unknown environments, characterized in that, The system includes: Perception module: Converts the unknown environment observation data collected by the robot at the current position into point cloud and aggregates it to generate the observed area. If the observed target is detected in the observed area, the position information and confidence level of the observed target are output. Imagination Module: Based on the observed area, a pre-trained generative model iteratively generates multiple unobserved areas from near to far. By counting the number of times the unobserved target appears in the multiple unobserved areas, the location and probability of the unobserved target are obtained. Decision Module: By integrating the location information and confidence level of observed targets, and the existence location and probability of unobserved targets, a comprehensive belief state reflecting a complete understanding of the preset environment is constructed. Based on the comprehensive belief state, the Imagine-POUCT algorithm is used to simulate all possible movement positions of the robot in the discrete action space. Each movement position is scored using a dense reward function, and the robot's highest-scoring movement position is determined. The Imagine-POUCT algorithm, based on the tree search strategy of the PUCT algorithm, recreates the search tree based on the new comprehensive belief state after each robot movement. Perception-Imagination-Decision Loop Module: Iterate through the above steps until a preset number of observed targets are successfully found, and the confidence level of these observed targets is greater than a preset threshold. Then stop iterating and output the location information of all observed targets.

6. A computer system, characterized in that, include: At least one processor; And a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the robot autonomous target search method for unknown environments as described in any one of claims 1 to 4.

7. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, is capable of performing the autonomous target search method for robots in unknown environments as defined in any one of claims 1 to 4.

Images