Unmanned sweeping robot navigation method based on multi-sensor fusion

Abstract

The invention discloses an unmanned sweeping robot navigation method based on multi-sensor fusion, and belongs to the technical field of sweeping robots. According to the method, an environment sensing system is constructed through multi-source sensor fusion, and efficient path planning and obstacle avoidance decision making in a dynamic environment are realized in combination with a quantum heuristic behavior decision engine; a traditional decision problem is converted into quantum state superposition operation through a quantum state probability amplitude model, and dynamic coupling of environment perception and a task target is achieved through the joint effect of a complex phase angle and a weight coefficient; and the dynamic collapse equation improves the adaptive capacity of behavior selection. The sweeping efficiency, safety and environment adaptability of the sweeping robot in a complex dynamic environment are remarkably improved, and the sweeping robot is suitable for various scenes such as families and offices.

Description

Technical Field

[0001] This invention relates to the field of robotic vacuum cleaner technology, and in particular to a navigation method for unmanned robotic vacuum cleaners based on multi-sensor fusion. Background Technology

[0002] Autonomous robotic vacuum cleaners are intelligent devices capable of automatically cleaning floors. They integrate environmental perception, path planning, autonomous navigation, and cleaning technologies, efficiently completing cleaning tasks without human intervention. However, current technologies suffer from several drawbacks. Traditional state machine decision-making frameworks are prone to getting stuck in local optima in dynamic scenarios, leading to path repetition and inefficiency. Single-positioning modes lack stability in narrow or obstructed environments, resulting in drift. Safety features rely on single sensor triggers, making it difficult to handle multi-dimensional dynamic obstacles. Furthermore, energy management lacks dynamic scheduling mechanisms, making it difficult to balance battery life and coverage.

[0003] Therefore, it is necessary to propose a comprehensive navigation method that integrates multi-source sensors and quantum heuristic decision-making models. Summary of the Invention

[0004] Purpose of the invention: To address the above problems, the purpose of this invention is to provide a navigation method for unmanned sweeping robots based on multi-sensor fusion.

[0005] Technical solution: The navigation method for an unmanned sweeping robot based on multi-sensor fusion of the present invention includes the following steps:

[0006] S1 deploys sensors on the unmanned sweeping robot, including radar, binocular vision camera, ultrasonic array, IMU unit and wheel encoder, and uses the sensors to collect multi-source sensor data.

[0007] S2, by fusing multi-source sensor data through an improved ResNet-18 model, obtains an environmental probability map containing semantic information;

[0008] S3 generates a dynamically occupied grid map, and updates spatial classification information, obstacle attributes and terrain information in real time on the dynamically occupied grid map;

[0009] S4 combines a dynamic grid map and switches between dual-mode positioning. In open areas, an improved NDT algorithm is used for positioning, while in narrow areas, fusion positioning is used to obtain the pose information of the robot vacuum cleaner.

[0010] S5 uses pose information as the starting point to generate a globally optimal path from the starting point to the end point, and performs local obstacle avoidance while maintaining the globally optimal path.

[0011] S6 decomposes the cleaning task into discrete quantum state behavior units. Based on the quantum-inspired optimized simulated annealing mechanism and dynamic collapse equation, the optimal action behavior sequence is obtained. Each quantum state behavior unit implements the corresponding behavior decision according to the optimal action behavior sequence.

[0012] S7 dynamically reconstructs the connection relationships of nodes in the topology map based on the positioning results and behavioral decision results;

[0013] S8 operates independently with a three-level security protection mechanism; if any mechanism is triggered, obstacle avoidance will be implemented.

[0014] Furthermore, S2 includes:

[0015] The input layer of the ResNet-18 model is changed to a dual-channel input layer. The asynchronously processed environment point cloud height map and RGB image stream are input into the dual-channel input layer, respectively. After passing through the convolutional layer, residual block layer, and spatial pyramid pooling layer in sequence, the environment probability map is output through the output layer to represent the semantic category probability distribution at each pixel location.

[0016] Furthermore, S3 includes:

[0017] Point cloud data is downsampled using voxel grids, then ground segmentation is performed, and outliers are removed to obtain effective detection point cloud. Based on the effective detection point cloud, the probability value of each grid cell being occupied by an obstacle is calculated.

[0018] Visual data is semantically segmented using a pre-trained DeepLabv3+ model to obtain obstacle attributes, including static and dynamic obstacles, and their corresponding confidence scores.

[0019] A static map is formed based on static obstacles, and a dynamic layer is formed based on dynamic obstacles. Based on the static map and the dynamic layer, the actual ground environment is transformed into a three-dimensional voxel grid. The spatial classification information of the three-dimensional voxel grid is divided into passable areas, obstacle areas and uncertain areas according to probability values.

[0020] The grid cells are used to record the occupancy of the corresponding height, and areas with a height greater than the threshold are marked as dangerous areas;

[0021] Special areas are marked on a three-dimensional voxel grid, including charging station locations, user-defined restricted areas, key cleaning areas, and boundary information;

[0022] The last update time is recorded for each grid cell, and the dynamically occupied grid map is updated when a new obstacle is detected.

[0023] Furthermore, S4 includes:

[0024] When the effective detection range of the lidar is below the detection threshold, the narrow area positioning mode is activated; otherwise, the open area positioning mode is activated.

[0025] Furthermore, the steps for localization in open areas using the improved NDT algorithm include:

[0026] The raw environmental point cloud data acquired by the LiDAR is preprocessed to separate ground points from non-ground points and remove outliers. Then, the initial pose estimation of the robot vacuum is obtained through the coarse registration stage, and the high-precision pose of the robot vacuum is obtained through the fine registration stage.

[0027] Furthermore, the steps for employing fusion localization in narrow areas include:

[0028] By integrating visual sensor data and data collected by wheel encoders through an extended Kalman filter, the real-time pose information of the robot vacuum cleaner is obtained, including position coordinates, orientation angle, timestamp, and confidence level.

[0029] Furthermore, S5 includes:

[0030] Using the starting point as the starting tree and the ending point as the ending tree, the bidirectional RRT* algorithm is used to iteratively search the two search trees. When the two trees are successfully connected, the complete path is extracted as the optimal path, and a path smoothing algorithm is applied for post-processing to obtain the globally optimal path.

[0031] Through model predictive control, dynamic obstacles are responded to in real time, and the safety boundary is adjusted in real time according to the safety distance and the speed of the dynamic obstacle.

[0032] Furthermore, the dynamic collapse equation is expressed as:

[0033] ,

[0034] in, This represents the selection probability of the j-th quantum state behavioral unit. This represents the utility value of the behavior after decoherence. This represents the simulated annealing coefficients based on quantum-inspired optimization. This represents the time interval parameter, which is inversely proportional to the robot's speed. [⋅] represents the matrix trace operation. Represents the density matrix, This represents a projection operator that maps specific behavioral units to the quantum state space.

[0035] Furthermore, based on the quantum-inspired optimized simulated annealing mechanism and dynamic collapse equation, the steps to obtain the optimal action sequence include:

[0036] S61, randomly select candidate behavior j from a predefined set of behavior units, which includes basic actions such as movement, vacuuming, and obstacle avoidance;

[0037] S62, calculate the energy difference between the candidate action and the current action, and quantify the cost of action switching by comparing the squared difference of the utility values ​​of the two action units;

[0038] S63 introduces the quantum tunneling effect to correct the traditional Metropolis criterion and calculates the quantum tunneling probability;

[0039] S64: When the energy difference is negative, the candidate behavior is accepted directly; otherwise, a random judgment is made according to the calculated quantum tunneling probability. If the random number is less than the quantum tunneling probability, the new behavior is adopted to achieve a jump from the local optimum.

[0040] S65, the current temperature decays according to the cooling rate after each iteration;

[0041] S66: S61 to S65 are executed repeatedly until the temperature drops to the termination threshold. After the algorithm finishes its iteration, it outputs the optimal action sequence after global optimization.

[0042] Furthermore, the steps for acquiring multi-source sensor data using sensors include:

[0043] LiDAR data is acquired using LiDAR, including environmental point cloud data, distance measurement data, reflection intensity, timestamps, scanning angles, and quality marker data.

[0044] Visual sensor data is acquired using a binocular vision camera, including RGB image streams, depth images, disparity maps, feature point data, optical flow data, semantic segmentation data, and exposure parameters.

[0045] The ultrasonic array is used to collect distance measurement data, echo time data, signal strength data, temperature compensation data, and blind zone detection data.

[0046] The IMU unit is used to measure triaxial acceleration, triaxial angular velocity, triaxial magnetic field, temperature, attitude angle, acceleration bias, and gyroscope bias data.

[0047] The wheel encoder is used to collect pulse count data, speed data, odometer data, direction indicator data, motor current data, slip detection data, and cumulative mileage data.

[0048] Beneficial effects: Compared with the prior art, the significant advantages of this invention are:

[0049] 1. In this invention, the quantum state probability amplitude model transforms the traditional decision-making problem into a quantum state superposition operation. Through the combined effect of complex phase angles and weighting coefficients, it achieves dynamic coupling between environmental perception and task objectives. The dynamic collapse equation is introduced from non-equilibrium thermodynamics. This feature enables the behavior selection process to have adaptive adjustment capabilities, and can quickly trigger secondary superposition calculations when sudden obstacles appear, achieving dual optimization of cleaning efficiency and safety in complex dynamic environments.

[0050] 2. This invention significantly improves the environmental adaptability of the cleaning robot through multi-sensor fusion and intelligent decision-making mechanisms: the collaborative work of LiDAR and visual cameras achieves millimeter-level obstacle detection accuracy, and combined with dynamic environment modeling algorithms, it can update the cleaning path in real time in scenarios where moving objects frequently appear; the innovative application of the quantum-heuristic behavior decision engine enables the robot to have multi-objective collaborative decision-making capabilities, and can quickly generate the optimal action sequence that balances efficiency and safety when faced with suddenly appearing obstacles or complex furniture layouts, reducing path repetition rate compared to traditional decision-making systems; the dual-mode positioning system ensures stable operation in different spatial environments, and can maintain centimeter-level positioning accuracy even in narrow areas where laser detection is limited.

[0051] 3. This invention effectively reduces the risk of falling down stairs or colliding with valuables through the dual protection of hardware emergency stop and dynamic obstacle prediction: the adaptive topology map update mechanism enables the robot to continuously adapt to changes in the home environment, and maintains a map matching accuracy of over 90% even after long-term use; combined with task scheduling of the intelligent energy consumption model, the battery life is increased by about 24% while ensuring daily cleaning coverage, making it particularly suitable for the continuous cleaning needs of large households. Attached Figure Description

[0052] Figure 1 This is a flowchart of the present invention;

[0053] Figure 2 This is an experimental comparison diagram between the present invention and the conventional solution. Detailed Implementation

[0054] The embodiments of the present invention will be further described in detail below with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are merely illustrative of the present invention and not intended to limit the scope of the invention. Furthermore, it should be noted that, for ease of description, the accompanying drawings show only the parts relevant to the embodiments of the present invention, and not all structures.

[0055] In the following description, specific details such as target system architecture and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods are omitted so as not to obscure the description of this application with unnecessary detail.

[0056] It should be understood that, when used in this application specification and the appended claims, the term "comprising" indicates the presence of the described features, integrals, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or a collection thereof.

[0057] It should also be understood that the term “and / or” as used in this application specification and the appended claims means any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0058] Furthermore, in the description of this application and the appended claims, the terms "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0059] References to "one embodiment" or "some embodiments" in this specification mean that one or more embodiments of this application include the target features, structures, or characteristics described in connection with that embodiment. Therefore, the phrases "in one embodiment," "in some embodiments," "in other embodiments," "in still other embodiments," etc., appearing in different parts of this specification do not necessarily refer to the same embodiment, but rather mean "one or more, but not all, embodiments," unless otherwise specifically emphasized.

[0060] The flowchart of the navigation method for an unmanned sweeping robot based on multi-sensor fusion described in this embodiment is as follows: Figure 1 As shown, it includes the following steps:

[0061] S1 deploys sensors on the unmanned sweeping robot, including radar, binocular vision camera, ultrasonic array, IMU unit and wheel encoder, and uses the sensors to collect multi-source sensor data.

[0062] Specifically, the steps for acquiring multi-source sensor data using sensors include:

[0063] LiDAR data is acquired using LiDAR, including environmental point cloud data, distance measurement data, reflection intensity, timestamps, scanning angles, and quality marker data.

[0064] Visual sensor data is acquired using a binocular vision camera, including RGB image streams, depth images, disparity maps, feature point data, optical flow data, semantic segmentation data, and exposure parameters.

[0065] The ultrasonic array is used to collect distance measurement data, echo time data, signal strength data, temperature compensation data, and blind zone detection data.

[0066] The IMU unit is used to measure triaxial acceleration, triaxial angular velocity, triaxial magnetic field, temperature, attitude angle, acceleration bias, and gyroscope bias data.

[0067] The wheel encoder is used to collect pulse count data, speed data, odometer data, direction indicator data, motor current data, slip detection data, and cumulative mileage data.

[0068] In one example, a LiDAR-Slam device is used as the core sensor, configured with a 360° rotating laser module scanning at a frequency of 6 scans / second, a detection radius of 8 meters, an angular resolution of 0.1°, a point cloud density of 2000 3D coordinate points (x, y, z) per second, and an accuracy of ±2cm. It detects the distance to obstacles within an 8-meter radius and collects the reflection intensity values ​​of the laser beam on different surfaces (0-255) to identify materials such as floors. Timestamp data refers to the precise time stamp (microsecond level) of each point cloud frame; scan angle data refers to the angular information of each measurement point during the 360° rotation; and quality marker data refers to the quality assessment value of each ranging point, indicating data reliability.

[0069] The quality assessment value is an evaluation index of the reliability of the data of each ranging point by the lidar. The specific evaluation methods include: (1) Signal strength assessment: based on the laser beam reflection intensity value (range 0-255), the intensity value is greater than 150 as high quality, 100-150 as medium quality, and less than 100 as low quality. (2) Multiple echo detection: detect whether there is multipath reflection: such as single echo: quality value is 1.0; multiple echo: quality value decreases by 0.8-0.3. (3) Distance consistency verification: gradient analysis of distance change with adjacent measurement points: gradient less than 5%: quality value 1.0; gradient between 5% and 15%: quality value 0.7-0.9; gradient greater than 15%: quality value less than 0.7. (4) Timestamp continuity: check the continuity of measurement timestamps: if the continuity is good, the quality weight is +0.1; if there is a jump, the quality weight is -0.2.

[0070] The quality assessment formula can be expressed as:

[0071] Quality_score = 0.4×Signal_strength + 0.3×Echo_consistency + 0.2×Distance_gradient + 0.1×Time_continuity,

[0072] Among them, Signal_strength represents the normalized score of laser beam reflection intensity, Echo_consistency represents the evaluation result of multiple echo detection, Distance_gradient represents the gradient analysis result of the distance change with adjacent measurement points, and Time_continuity represents the evaluation of the continuity of measurement timestamps. These four parameters are weighted and summed (with weights of 0.4, 0.3, 0.2, and 0.1, respectively) to comprehensively evaluate the data quality of each laser ranging point, and finally obtain a quality score in the range of 0-1, which is used for reliability judgment in subsequent data fusion and environmental mapping.

[0073] Visual sensor data, including RGB image streams, depth images, disparity maps, feature point data, optical flow data, semantic segmentation data, and exposure parameters, is acquired using a binocular vision camera. The binocular vision camera uses a global shutter CMOS sensor with a baseline length of 12cm, a resolution of 1080p, a frame rate of 30FPS, and supports stereo matching for depth calculation. The acquired RGB image stream refers to a sequence of color images at 1080p resolution and 30FPS. Pixel-level depth data (0-5 meters) calculated using a stereo matching algorithm is used as the depth image. The left and right views acquired by the 12cm baseline binocular camera are used as disparity maps, with the pixel difference matrix between the left and right images used for depth calculation. A global shutter timestamp is acquired to ensure time synchronization of image acquisition. Feature point data refers to ORB feature point coordinates and descriptors (approximately 500-1000 per frame). Optical flow data refers to pixel motion vectors between consecutive frames, used for dynamic object detection. Semantic segmentation data refers to pixel-level annotation information for 20 categories of household objects. Exposure parameters refer to the aperture, shutter speed, and ISO values ​​for automatic exposure control.

[0074] The system utilizes an ultrasonic array to collect distance measurement data, echo time data, signal strength data, temperature compensation data, and blind spot detection data. The ultrasonic array consists of eight 40kHz transceiver modules, with a detection range of 5cm to 3m, covering the blind spot under the robot vacuum. Distance measurement data refers to distance values ​​in eight directions (5cm-3m), with an accuracy of ±1cm; echo time data refers to the round-trip time of the 40kHz sound wave (microseconds); signal strength data refers to the amplitude of the echo signal, used to determine surface material; temperature compensation data refers to the correction value for the effect of ambient temperature on the speed of sound; and blind spot detection data provides information on obstacles less than 5cm below the robot's bottom. The ultrasonic array can also detect multipath effect markers to identify false echoes caused by multiple reflections.

[0075] The IMU unit measures triaxial acceleration, triaxial angular velocity, triaxial magnetic field, temperature, attitude angle, acceleration bias, and gyroscope bias data. The IMU unit integrates a triaxial accelerometer, gyroscope, and magnetometer, with a sampling frequency of 200Hz, enabling attitude calculation and motion compensation. The triaxial acceleration data refers to the X, Y, and Z-axis acceleration values ​​(±8g range); the triaxial angular velocity data refers to the X, Y, and Z-axis angular velocity values ​​(±2000° / s range); the triaxial magnetic field data refers to the geomagnetic field strength vector, used for magnetic heading calculation; the temperature data refers to the sensor's operating temperature, used for temperature drift compensation; the attitude angle data refers to the filtered pitch, roll, and yaw angles; the acceleration bias data refers to static deviation calibration parameters; and the gyroscope bias data refers to zero-point drift calibration parameters. All data are updated at a frequency of 200Hz (every 5ms).

[0076] Wheel encoders are used to collect pulse count data, rotational speed data, odometer data, direction marking data, motor current data, slip detection data, and cumulative mileage data. The wheel encoders employ photoelectric incremental sensors, producing 500 pulses per revolution, combined with motor current feedback to achieve millimeter-level odometer accuracy. The pulse count data refers to the number of pulses from the left and right wheel encoders (500 pulses per revolution); the rotational speed data refers to the real-time wheel speed (revolutions per minute) with an accuracy of 0.1 rpm; the odometer data refers to the distance traveled based on the wheel diameter (millimeter-level accuracy); the direction marking data refers to the forward / backward / left / right turn status indicators; the motor current data refers to the operating current of the left and right motors, reflecting the load condition; the slip detection data refers to the detection of tire slippage through the ratio of current to rotational speed; and the cumulative mileage data is the total distance traveled by the robot.

[0077] All sensors achieve microsecond-level time synchronization via hardware triggering. The IEEE 1588 precision clock protocol provides a microsecond-level synchronization time scale. Spatial calibration adopts a joint calibration method using a checkerboard target and a laser reflector to ensure that all sensor data are precisely aligned in the time dimension.

[0078] S2 uses an improved ResNet-18 model to fuse multi-source sensor data to obtain an environmental probability map containing semantic information.

[0079] Furthermore, S2 includes:

[0080] The input layer of the ResNet-18 model is changed to a dual-channel input layer. The asynchronously processed environment point cloud height map and RGB image stream are input into the dual-channel input layer, respectively. After passing through the convolutional layer, residual block layer, and spatial pyramid pooling layer in sequence, the environment probability map is output through the output layer to represent the semantic category probability distribution at each pixel location.

[0081] Table 1 shows a structural comparison between the original ResNet-18 model and the improved ResNet-18 model. The specific improvements to the ResNet-18 model in this example include the input layer, residual blocks, and pooling strategy. The original input layer is a single-channel RGB image with a size requirement of [Batch, 3, 224, 224], where Batch represents the batch size, typically set to 8-32, 3 represents the three RGB color channels, and 224×224 represents the image resolution. The improved input layer is dual-channel, with one channel for the laser height map and the other for the RGB image, with a size requirement of [Batch, 4, 256, 256]. Of the four channels, one is the laser height map, and the other three are RGB images. Regarding the improvement of the residual blocks, dilated convolutions with dilation rates of 2 and 4 are used in the 3rd and 4th residual blocks, and a spatial attention module (SAM) is added after each residual block. A cross-modal feature fusion layer is added after the 2nd and 4th layers. The original ResNet-18 model only has downsampling paths, while the improved U-Net-style encoder-decoder structure uses an improved ResNet-18 backbone for the encoder and four layers of transposed convolutional upsampling for the decoder. Skip connections are used to connect the encoder-decoder layers at the corresponding resolutions. In addition, the original global average pooling is replaced with spatial pyramid pooling, and finally, a semantic probability map is output through the output layer.

[0082] Table 1

[0083] Before the environmental point cloud height map and RGB image stream are input to the dual-channel input layer, asynchronous processing is required due to their asynchronous acquisition characteristics in time and space. The asynchronous characteristics of the LiDAR data include: a scanning frequency of 6 scans / second (approximately 166.7ms per frame), data format of 360° rotated point cloud data, 2000 points per second; and a time characteristic where mechanical rotation results in different timestamps for each point. The asynchronous characteristics of the vision sensor data include: a frame rate of 30FPS (33.3ms per frame), data format of 1080p resolution RGB images and depth information, and a time characteristic of global shutter CMOS, synchronized across the entire frame but asynchronous with the LiDAR. Because the sampling times of the LiDAR and camera are mismatched, resulting in time asynchrony, the timestamps of each sensor are synchronized via the IEEE 1588 protocol, with a time deviation of less than 1 microsecond. Spatially, the coordinate systems of different sensors are aligned to the robot's body coordinate system, and the calibration error is controlled within ±2mm. The environmental point cloud acquired by the LiDAR is transformed from the sensor coordinate system to the robot's body coordinate system. The RANSAC algorithm is used to fit the ground plane, separating ground points from non-ground points. The height value of each non-ground point relative to the ground is calculated, and then projected to generate a 2D height map. This height map is input into the dual-channel input layer. The height map has a resolution of 5cm × 5cm grid precision, a size of 256 × 256 pixels (corresponding to an actual range of 12.8m × 12.8m), a value range of 0-2.5 meters (normalized to 0-255 grayscale values), and an update frequency of 6Hz (synchronized with the LiDAR). RGB images are acquired using the left camera of the binocular vision camera, with a resolution of 1080p (1920 × 1080), a standard RGB color space, and a sampling frequency of 30FPS. The RGB image was resized from 1920×1080 to 256×256, and color correction was performed, including white balance and gamma correction. Noise reduction was then applied using Gaussian filtering. Normalization was performed to normalize pixel values ​​to the [0,1] range. Data augmentation was then performed, including random rotation, flipping, and brightness adjustment. The image was categorized into 20 types of household objects, such as tables, chairs, sofas, and pets.

[0084] The environment point cloud heightmap and RGB image stream are input into the improved ResNet-18 model, where the environment point cloud heightmap is represented as:

[0085] ,

[0086] in, A 2D height map of the environmental point cloud projection; Next month

[0087] RGB image stream is represented as:

[0088] ,

[0089] in, This represents standardized preprocessed color image data with dimensions [batch, 3, 256, 256]. It includes the following channels: R channel (representing the red component [0, 1]), G channel (representing the green component [0, 1]), and B channel (representing the blue component [0, 1]). The physical coverage area is 12.8m × 12.8m, with each pixel corresponding to a 5cm × 5cm physical area. The semantic information includes the color, texture, and shape features of 20 types of home furnishing objects. The temporal characteristics are 30FPS acquisition, synchronized with laser data.

[0090] The features obtained by weighted fusion are represented as follows:

[0091] ,

[0092] in, This represents the weighting coefficient for laser data; for example, setting it to 0.7. This represents the weighting coefficient for visual data, such as setting it to 0.3.

[0093] The environmental semantic probabilistic graph outputs pixel-level semantic probabilities, represented as follows:

[0094] ,

[0095] in, This represents the activation function. This represents the semantic label of the i-th class.

[0096] The output dimension of the environmental semantic probabilistic map is [20, 256, 256]. 20 corresponds to 20 semantic categories, including sofa, table, chair, bed, cabinet, TV, refrigerator, washing machine, plant, trash can, shoes, toy, pet, person, door, window, carpet, table lamp, decoration, and others; 256×256 corresponds to a physical space of 12.8m×12.8m; the numerical meaning represents the category probability distribution of each pixel location, with a probability range of [0, 1], which is normalized by the Softmax function.

[0097] In this step, multi-source data fusion is achieved through an improved ResNet-18 two-stream architecture, realizing deep fusion of laser point cloud height maps and visual RGB images, ultimately outputting an environmental probability map containing 20 classes of semantic information. The entire system uses a dynamic weight adjustment mechanism to adaptively optimize the fusion results based on the real-time confidence of the sensors, providing a high-quality environmental perception foundation for subsequent path planning and decision-making.

[0098] S3 generates a dynamically occupied grid map, and updates spatial classification information, obstacle attributes, and terrain information in real time on the dynamically occupied grid map.

[0099] Furthermore, S3 includes:

[0100] Point cloud data is downsampled using voxel grids for noise reduction. Then, the RANSAC algorithm is used to segment the ground and fit the ground normal vector, and outliers are removed to obtain effective detection point cloud. Based on the effective detection point cloud, the probability value of each grid cell corresponding to the location is occupied by an obstacle is calculated, with the probability value ranging from [0,1].

[0101] Visual data is semantically segmented using a trained DeepLabv3+ model to obtain obstacle attributes, including static and dynamic obstacles, and their corresponding confidence scores; static obstacles include objects such as walls and furniture, while dynamic obstacles include objects such as pets and people.

[0102] A static map is formed based on static obstacles, and a dynamic layer is formed based on dynamic obstacles. Based on the static map and the dynamic layer, the actual ground environment is transformed into a three-dimensional voxel grid. The spatial classification information of the three-dimensional voxel grid is divided into passable areas, obstacle areas and uncertain areas according to probability values.

[0103] The grid cells are used to record the occupancy of the corresponding height, and areas with a height greater than a threshold (such as 3cm) are marked as dangerous areas;

[0104] Special areas are marked on a three-dimensional voxel grid, including charging station locations, user-defined restricted areas, key cleaning areas, and boundary information;

[0105] The last update time is recorded for each grid cell, and the dynamically occupied grid map is updated when a new obstacle is detected.

[0106] In one example, for a voxel cell (i,j,k), the formula for the probability that the position of each grid cell is occupied by an obstacle is:

[0107] ,

[0108] in, This represents the posterior probability of occupancy based on sensor data. Indicates the likelihood of sensor observations. This represents the prior probability of occupancy. Represents the marginal probability of sensor data;

[0109] The threshold is set to 0.65. If the calculated occupancy probability value is greater than 0.65, it is determined that an obstacle exists.

[0110] In one example, spaces with a probability value less than 0.35 are designated as passable areas, through which the robot can safely pass; spaces with a probability value greater than 0.65 are designated as obstacle areas and marked as impassable; and spaces with a probability value between 0.35 and 0.65 are designated as uncertain areas, which the robot must handle with caution.

[0111] For special area markings, the charging station location is a fixed reference point, the prohibited area is the user-defined area that will not be cleaned, the key cleaning area is the area that needs to be cleaned more thoroughly, and the boundary information is the structural information such as room boundaries and corners.

[0112] The dynamically constructed grid map provides rich environmental information for subsequent localization and path planning, enabling the robot to achieve precise navigation and efficient cleaning in complex dynamic environments.

[0113] S4 combines a dynamically occupied grid map and switches between dual-mode positioning. In open areas, an improved NDT algorithm is used for positioning, while in narrow areas, fusion positioning is used to obtain the pose information of the robot vacuum cleaner.

[0114] Furthermore, S4 includes:

[0115] When the effective detection range of the lidar is below the detection threshold, the narrow area positioning mode is activated; otherwise, the open area positioning mode is activated.

[0116] Furthermore, the steps for localization in open areas using the improved NDT algorithm include:

[0117] The raw environmental point cloud data acquired by the LiDAR is preprocessed to separate ground points from non-ground points and remove outliers. Then, the initial pose estimation of the robot vacuum is obtained through the coarse registration stage, and the high-precision pose of the robot vacuum is obtained through the fine registration stage.

[0118] In one example, during the coarse registration stage, the lidar point cloud is divided into a 40cm × 40cm grid. The point cloud within each grid is approximated by a Gaussian distribution. The mean μ and covariance matrix Σ of the grid center points are calculated, and the transformation matrix T is solved using Newton's method to minimize the objective function. ,

[0119] The rough pose transformation matrix T is obtained by iterating 5-10 times;

[0120] Where p represents the coordinates of a single 3D point in the lidar point cloud, expressed as p = [x, y, z]. T .

[0121] Based on coarse registration, high-precision pose is obtained through a fine registration stage (10cm grid). The implementation process is as follows: using the coarse registration result as the initial value, a new 10cm×10cm fine grid is divided. Gaussian distribution parameters are recalculated within each small grid, and more stringent convergence conditions are used, such as matching error <3cm. Millimeter-level accuracy is achieved through 15-20 iterations.

[0122] Coarse registration reduces computation by 70% and improves overall speed by about 3 times. Fine registration ensures that the final error is controlled within 3cm. The two-stage strategy avoids getting trapped in local optima.

[0123] Before NDT registration, the laser point cloud is preprocessed. The RANSAC algorithm is used for ground segmentation to separate ground points from non-ground points, reducing ground interference and ensuring consistency of random sampling. Three points are randomly selected to fit the plane equation: ax + by + cz + d = 0. The distances from all points to the plane are calculated, and the number of interior points within a distance threshold (typically 5cm) is counted. This process is iterated 1000 times, and the plane with the most interior points is selected as the ground. Ground points are marked and removed from the registration point cloud. Outlier removal is performed using a statistical filter with the following parameters: neighbor number K = 50, standard deviation factor σ = 2.0, to remove noise points and outlier measurements, improving registration stability.

[0124] Fusion positioning is suitable for narrow areas. The specific triggering conditions are: the effective detection range of the lidar is less than 2.5 meters. Typical scenarios include corridors, gaps between furniture, and transition areas at doorways.

[0125] Furthermore, the steps for employing fusion localization in narrow areas include:

[0126] By integrating visual sensor data and data collected by wheel encoders through an extended Kalman filter, the real-time pose information of the robot vacuum cleaner is obtained, including position coordinates, orientation angle, timestamp, and confidence level.

[0127] The differences between fused positioning and open area positioning are shown in Table 2 below.

[0128] Table 2

[0129] The robot vacuum cleaner uses a dual-mode positioning method to obtain its pose information, including position coordinates (the robot's position in the global coordinate system), orientation angle (the robot's heading angle), timestamp (to ensure time synchronization with other modules), and confidence score (a reliability rating for pose estimation).

[0130] The obtained pose information can be used for map optimization, such as pose drift detection for difference analysis between consecutive poses, and loop closure detection trigger for matching degree between current pose and historical trajectory; it can also be used for environmental mapping in step S3, such as point cloud registration benchmark to unify sensor data to global coordinate system; occupy grid update to update map information based on accurate pose; it can also be supplied to step S5 for path planning, such as obtaining the current state as the starting point of the path planning algorithm, and motion control, with real-time pose feedback for motion trajectory tracking.

[0131] S5 uses pose information as the starting point to generate a globally optimal path from the starting point to the end point, and performs local obstacle avoidance while maintaining the globally optimal path.

[0132] Furthermore, S5 includes:

[0133] Using the starting point as the starting tree and the ending point as the ending tree, the bidirectional RRT* algorithm is used to iteratively search the two search trees. When the two trees are successfully connected, the complete path is extracted as the optimal path, and a path smoothing algorithm is applied for post-processing to obtain the globally optimal path.

[0134] Through model predictive control, dynamic obstacles are responded to in real time, and the safety boundary is adjusted in real time according to the safety distance and the speed of the dynamic obstacle.

[0135] Local obstacle avoidance is achieved using model predictive control, with a prediction time of 2 seconds, a control time of 0.5 seconds, and a sampling interval of 0.1 seconds.

[0136] S6 decomposes the cleaning task into discrete quantum state behavior units. Based on the quantum-inspired optimized simulated annealing mechanism and dynamic collapse equation, the optimal action behavior sequence is obtained. Each quantum state behavior unit implements the corresponding behavior decision according to the optimal action behavior sequence.

[0137] Furthermore, the dynamic collapse equation is expressed as:

[0138] ,

[0139] in, This represents the selection probability of the j-th quantum state behavioral unit. This represents the utility value of the behavior after decoherence. This represents the simulated annealing coefficients based on quantum-inspired optimization. This represents the time interval parameter, which is inversely proportional to the robot's speed. [⋅] represents the matrix trace operation. Represents the density matrix, This represents a projection operator that maps specific behavioral units to the quantum state space.

[0140] Furthermore, based on the quantum-inspired optimized simulated annealing mechanism and dynamic collapse equation, the steps to obtain the optimal action sequence include:

[0141] S61, randomly select candidate behavior j from a predefined set of behavior units, which includes basic actions such as movement, vacuuming, and obstacle avoidance;

[0142] S62, calculate the energy difference between the candidate action and the current action, and quantify the cost of action switching by comparing the squared difference of the utility values ​​of the two action units;

[0143] S63 introduces the quantum tunneling effect to correct the traditional Metropolis criterion and calculates the quantum tunneling probability;

[0144] S64: When the energy difference is negative, the candidate behavior is accepted directly; otherwise, a random judgment is made according to the calculated quantum tunneling probability. If the random number is less than the quantum tunneling probability, the new behavior is adopted to achieve a jump from the local optimum.

[0145] S65, the current temperature decays according to the cooling rate after each iteration;

[0146] S66: S61 to S65 are executed repeatedly until the temperature drops to the termination threshold. After the algorithm finishes its iteration, it outputs the optimal action sequence after global optimization.

[0147] The quantum tunneling effect is introduced to modify the traditional Metropolis criterion, and the quantum tunneling probability is calculated. The formula for the traditional Metropolis criterion is as follows:

[0148] ,

[0149] in, This represents the probability of accepting the new state. Indicates the current temperature parameter; The energy difference is expressed by the following formula:

[0150] ,

[0151] in, Let represent the squared utility value of candidate action j. This represents the squared utility value of the current action.

[0152] The core formula for quantum tunneling probability is:

[0153] ,

[0154] in, This represents the acceptance probability after quantum tunneling correction. This represents the quantum tunneling strength coefficient, with a value ranging from 0.1 to 0.5. This represents the reduced Planck constant, normalized to 1; This represents the Boltzmann constant (normalized to 1); Indicates the current annealing temperature; The characteristic frequency parameter is calculated using the following formula:

[0155] ,

[0156] in, This indicates the robot's current speed. This represents the characteristic scale of a typical obstacle and can be set to 0.3m.

[0157] The dynamic adjustment formula for the quantum tunneling strength coefficient is:

[0158] ,

[0159] in, This represents the basic tunnel penetration strength, which can be set to 0.2; This represents the environmental sensitivity coefficient, which can be set to 0.5. This indicates the number of obstacles currently detected. This indicates the historical average number of obstacles.

[0160] The temperature decay formula is:

[0161] ,

[0162] Where α represents the cooling rate, typically 0.95; initial temperature The termination temperature threshold is adaptively calculated based on the maximum behavioral utility value. .

[0163] when When, directly accept candidate actions; when When generating random numbers r∈[0,1], if If the candidate behavior is accepted, quantum tunneling is achieved; otherwise, the current behavior is maintained.

[0164] In the simulated annealing stage based on quantum-inspired optimization, the system controls the traversal range of the decision space using a temperature parameter, eliminating local optima traps through a gradual cooling process. The dynamic collapse equation, by introducing a task priority tensor and an environmental complexity factor, establishes a non-Hermitian Hamiltonian model, enabling the behavior selection process to adaptively adjust. When encountering sudden obstacles, the system triggers a double superposition operation, preserving the quantum entangled states of suboptimal behavior sequences as emergency decision reserves. This architecture overcomes the rigid switching defects of traditional state machines, significantly improving decision robustness in complex scenarios while ensuring real-time response speed.

[0165] S7 dynamically reconstructs the connection relationships of nodes in the topology map based on the positioning results and behavioral decision results.

[0166] Topology maps are high-level map representations independent of dynamic occupied grid maps. The difference between topology maps and occupied grid maps is that occupied grid maps are dense representations based on pixels / voxels, recording the occupied state of each location; topology maps are sparse representations based on graph theory, focusing only on key locations and connections. Dynamic occupied grid maps are used for local path planning and obstacle avoidance, while topology maps are used for global task scheduling and room-level path planning.

[0167] In the topology map, nodes include room nodes (such as the center point of functional areas like the living room, bedroom, and kitchen), doorway nodes (such as transitional passages between rooms), charging station nodes (such as charging station locations), key cleaning points (such as areas requiring intensive cleaning: under the bed, under the sofa), and navigation path points (such as key navigation points in a long corridor). The connections between nodes include reachability connections, weight information, and time attributes. Reachability connections indicate that the robot can directly reach another node from one node. Weight information includes distance, cleaning priority, and passage difficulty. Time attributes record the effective time of the connection (e.g., some doors may be closed).

[0168] Dynamically reconstruct nodes and connections in the topology map, including: adding nodes (such as when a new room or important location is discovered), deleting nodes (such as locations that have been inaccessible for a long time), updating weights (adjusting connection weights based on actual cleaning results), and optimizing topology (simplifying redundant connections and optimizing the graph structure).

[0169] The dynamic adjustment of map nodes is based on a graph optimization framework, using the g2o library for pose graph optimization. Node connections are updated based on the confidence level of laser loop closure detection, with a loop closure matching score threshold of 0.75. The incremental mapping module detects the boundaries of unexplored areas in real time. When a new obstacle is detected by ultrasonic waves and persists for more than 10 seconds, the map expansion process is triggered. The topology optimization frequency is 1Hz, supporting the annotation of furniture or temporary obstacles that suddenly appear in unstructured environments. The node spacing is adaptively adjusted from 30cm to 1.5m.

[0170] S8 operates independently with a three-level security protection mechanism; if any mechanism is triggered, obstacle avoidance will be implemented.

[0171] The hardware emergency stop loop has a response time of less than 50ms, detects collision acceleration using a Hall effect sensor, and has a trigger threshold of 2g. Dynamic obstacle prediction is based on the YOLOv5s model, with an input resolution of 640×640, an inference speed of 30FPS, and supports the recognition of 20 common household items. The ultrasonic fall avoidance system uses dual-frequency pulse detection, alternating between 40kHz and 25kHz transmission frequencies, and triggers emergency braking when the height difference in the detection blind zone exceeds 3cm. The safety system employs a redundant design, with three levels of protection operating independently; any condition triggered will immediately execute the obstacle avoidance strategy.

[0172] Collect energy consumption data and map coverage indicators of the sweeping device system's behavioral units, dynamically adjust motor control parameters, and optimize task scheduling priorities in path planning.

[0173] The battery management system monitors individual cell voltages in real time with a sampling accuracy of ±1%, dynamically adjusts the motor PWM duty cycle, and controls the speed within a range of 0.1~0.8 m / s. The task scheduler is based on an improved EDF (Earliest Deadline First) algorithm, with priority weights allocated according to area cleanliness (85%), remaining battery power (10%), and task urgency (5%). The mathematical model of the improved EDF algorithm is as follows:

[0174] ,

[0175] The values ​​of each weight coefficient are divided into the following categories: =0.85, =0.10, =0.05, Indicates the cleanliness of the area. Indicates the remaining battery power. Indicates the urgency of the task.

[0176] Motion parameter optimization employed a genetic algorithm with a population size of 50 and 100 iterations. Optimization variables included linear acceleration (0.2~0.5 m / s²) and angular velocity (0.3~1.2 rad / s). The endurance balance point was determined through Pareto front analysis, with a sweep coverage to energy consumption weighting ratio of 2:1.

[0177] To further illustrate the effectiveness and superiority of the unmanned sweeping robot navigation method described in this invention, the following examples are provided.

[0178] The comparison uses a traditional industry solution with the following parameters:

[0179] The decision-making system employs a finite state machine (FSM), with the following state transition rules: obstacle avoidance when the obstacle distance is <0.3m, and recharging when the battery level is <20%. Localization and navigation utilize the SLAM algorithm, with path planning using the A* (A-star Algorithm) for global planning and the DWA (Dynamic Window Approach) for local obstacle avoidance. The safe distance is set to a fixed 0.4m, and energy management is set to a constant motor speed of 0.3m / s. The A* algorithm provides a globally optimal path as a reference line, while the DWA algorithm makes local adjustments based on real-time environmental information. When encountering dynamic obstacles, the DWA algorithm temporarily deviates from the A* path; after the obstacle disappears, the DWA algorithm guides the robot back to the A* path.

[0180] First, we compared the path planning efficiency by repeating the test 20 times in a living room scene with 4 dynamic obstacles. The results are shown in Table 3.

[0181] Table 3

[0182]

[0183] Next, the positioning accuracy was compared. The system was run continuously for 10 minutes in a 60cm wide laser-blocked channel. The results are shown in Table 4 below.

[0184] Table 4

[0185]

[0186] Then, a dynamic obstacle avoidance performance test was conducted, simulating a pet suddenly approaching (relative speed 0.8 m / s), and the results are shown in Table 5.

[0187] Table 5

[0188]

[0189] Finally, the energy consumption optimization effect was compared. Under the condition of a battery capacity of 3000mAh, the comparison results are shown in Table 6.

[0190] Table 6

[0191]

[0192] Figure 2 The comparison between the present invention and traditional methods in several indicators, including path repetition rate, cleaning time, cumulative drift error, number of positioning failures, average response delay, obstacle avoidance success rate, battery life, and energy consumption per unit area, shows that the present invention outperforms traditional methods in all indicators, significantly improving the cleaning efficiency, safety, and environmental adaptability of robotic vacuum cleaners in complex and dynamic environments.

[0193] This invention significantly enhances the environmental adaptability of the cleaning robot through multi-sensor fusion and intelligent decision-making mechanisms. The collaborative operation of LiDAR and a vision camera achieves millimeter-level obstacle detection accuracy, and combined with dynamic environment modeling algorithms, it can update the cleaning path in real time in scenarios with frequent moving objects. The innovative application of a quantum-heuristic behavioral decision engine enables the robot to make multi-objective collaborative decisions. When faced with suddenly appearing obstacles or complex furniture layouts, it can quickly generate an optimal action sequence that balances efficiency and safety, reducing path repetition compared to traditional decision-making systems. A dual-mode positioning system ensures stable operation in different spatial environments, maintaining centimeter-level positioning accuracy even in narrow areas where laser detection is limited.

[0194] This invention effectively reduces the risk of falling down stairs or colliding with valuables through dual safeguards of hardware emergency stop and dynamic obstacle prediction. The adaptive topology map update mechanism allows the robot to continuously adapt to changes in the home environment, maintaining over 90% map matching accuracy even after long-term use. Combined with task scheduling using an intelligent energy consumption model, battery life is increased by approximately 25% while ensuring daily cleaning coverage, making it particularly suitable for the continuous cleaning needs of large homes. The superposition state operation characteristics of the quantum decision engine also significantly reduce computational resource consumption in complex decision-making scenarios, enabling low-power embedded devices to smoothly run advanced navigation algorithms.

Claims

1. A method for navigation of an unmanned floor-sweeping robot based on multi-sensor fusion, the method comprising: Includes the following steps: S1 deploys sensors on the unmanned sweeping robot, including radar, binocular vision camera, ultrasonic array, IMU unit and wheel encoder, and uses the sensors to collect multi-source sensor data. S2, by fusing multi-source sensor data through an improved ResNet-18 model, obtains an environmental probability map containing semantic information; S3 generates a dynamically occupied grid map, and updates spatial classification information, obstacle attributes and terrain information in real time on the dynamically occupied grid map; S4 combines a dynamic grid map and switches between dual-mode positioning. In open areas, an improved NDT algorithm is used for positioning, while in narrow areas, fusion positioning is used to obtain the pose information of the robot vacuum cleaner. S5 uses pose information as the starting point to generate a globally optimal path from the starting point to the end point, and performs local obstacle avoidance while maintaining the globally optimal path. S6 decomposes the cleaning task into discrete quantum state behavior units. Based on the quantum-inspired optimized simulated annealing mechanism and dynamic collapse equation, the optimal action behavior sequence is obtained. Each quantum state behavior unit implements the corresponding behavior decision according to the optimal action behavior sequence. S7 dynamically reconstructs the connection relationships of nodes in the topology map based on the positioning results and behavioral decision results; S8 operates independently with a three-level security protection mechanism; if any mechanism is triggered, obstacle avoidance will be implemented.

2. The multi-sensor fusion based navigation method for unmanned floor cleaning robots according to claim 1, wherein S2 include: The input layer of the ResNet-18 model is changed to a dual-channel input layer. The asynchronously processed environment point cloud height map and RGB image stream are input into the dual-channel input layer, respectively. After passing through the convolutional layer, residual block layer, and spatial pyramid pooling layer in sequence, the environment probability map is output through the output layer to represent the semantic category probability distribution at each pixel location. 3.The multi-sensor fusion based navigation method of the unmanned floor cleaning robot according to claim 2, wherein, S3 includes: Point cloud data is downsampled using voxel grids, then ground segmentation is performed, and outliers are removed to obtain effective detection point cloud. Based on the effective detection point cloud, the probability value of each grid cell being occupied by an obstacle is calculated. Visual data is semantically segmented using a pre-trained DeepLabv3+ model to obtain obstacle attributes, including static and dynamic obstacles, and their corresponding confidence scores. A static map is formed based on static obstacles, and a dynamic layer is formed based on dynamic obstacles. Based on the static map and the dynamic layer, the actual ground environment is transformed into a three-dimensional voxel grid. The spatial classification information of the three-dimensional voxel grid is divided into passable areas, obstacle areas and uncertain areas according to probability values. The grid cells are used to record the occupancy of the corresponding height, and areas with a height greater than the threshold are marked as dangerous areas; Special areas are marked on a three-dimensional voxel grid, including charging station locations, user-defined restricted areas, key cleaning areas, and boundary information; The last update time is recorded for each grid cell, and the dynamically occupied grid map is updated when a new obstacle is detected.

4. The multi-sensor fusion based navigation method for unmanned floor cleaning robots according to claim 3, wherein S4 include: When the effective detection range of the lidar is below the detection threshold, the narrow area positioning mode is activated. Otherwise, use the open area positioning mode.

5. The navigation method for an unmanned sweeping robot based on multi-sensor fusion according to claim 4, characterized in that, The steps for localization in open areas using the improved NDT algorithm include: The raw environmental point cloud data acquired by the LiDAR is preprocessed to separate ground points from non-ground points and remove outliers. Then, the initial pose estimation of the robot vacuum is obtained through the coarse registration stage, and the high-precision pose of the robot vacuum is obtained through the fine registration stage.

6. The navigation method for an unmanned sweeping robot based on multi-sensor fusion according to claim 5, characterized in that, The steps for using fusion localization in narrow areas include: By integrating visual sensor data and data collected by wheel encoders through an extended Kalman filter, the real-time pose information of the robot vacuum cleaner is obtained, including position coordinates, orientation angle, timestamp, and confidence level.

7. The navigation method for an unmanned sweeping robot based on multi-sensor fusion according to claim 6, characterized in that, S5 include: Using the starting point as the starting tree and the ending point as the ending tree, the bidirectional RRT* algorithm is used to iteratively search the two search trees. When the two trees are successfully connected, the complete path is extracted as the optimal path, and a path smoothing algorithm is applied for post-processing to obtain the globally optimal path. Through model predictive control, dynamic obstacles are responded to in real time, and the safety boundary is adjusted in real time according to the safety distance and the speed of the dynamic obstacle.

8. The navigation method for an unmanned sweeping robot based on multi-sensor fusion according to claim 7, characterized in that, The dynamic collapse equation is expressed as: ， in, This represents the selection probability of the j-th quantum state behavioral unit. This represents the utility value of the behavior after decoherence. This represents the simulated annealing coefficients based on quantum-inspired optimization. This represents the time interval parameter, which is inversely proportional to the robot's speed. [⋅] represents the matrix trace operation. Represents the density matrix, This represents a projection operator that maps specific behavioral units to the quantum state space.

9. The navigation method for an unmanned sweeping robot based on multi-sensor fusion according to claim 8, characterized in that, The steps to obtain the optimal action sequence based on the quantum-inspired optimized simulated annealing mechanism and dynamic collapse equation include: S61, randomly select candidate behavior j from a predefined set of behavior units, which includes basic actions such as movement, vacuuming, and obstacle avoidance; S62, calculate the energy difference between the candidate action and the current action, and quantify the cost of action switching by comparing the squared difference of the utility values ​​of the two action units; S63 introduces the quantum tunneling effect to correct the traditional Metropolis criterion and calculates the quantum tunneling probability; S64: When the energy difference is negative, the candidate behavior is accepted directly; otherwise, a random judgment is made according to the calculated quantum tunneling probability. If the random number is less than the quantum tunneling probability, the new behavior is adopted to achieve a jump from the local optimum. S65, the current temperature decays according to the cooling rate after each iteration; S66: S61 to S65 are executed repeatedly until the temperature drops to the termination threshold. After the algorithm finishes its iteration, it outputs the optimal action sequence after global optimization.

10. The navigation method for an unmanned sweeping robot based on multi-sensor fusion according to any one of claims 1 to 9, characterized in that, The steps for acquiring multi-source sensor data using sensors include: LiDAR data is acquired using LiDAR, including environmental point cloud data, distance measurement data, reflection intensity, timestamps, scanning angles, and quality marker data. Visual sensor data is acquired using a binocular vision camera, including RGB image streams, depth images, disparity maps, feature point data, optical flow data, semantic segmentation data, and exposure parameters. The ultrasonic array is used to collect distance measurement data, echo time data, signal strength data, temperature compensation data, and blind zone detection data. The IMU unit is used to measure triaxial acceleration, triaxial angular velocity, triaxial magnetic field, temperature, attitude angle, acceleration bias, and gyroscope bias data. The wheel encoder is used to collect pulse count data, speed data, odometer data, direction indicator data, motor current data, slip detection data, and cumulative mileage data.

Images