Mobile robot control device giving consideration to field navigation patrol and transportation load

Abstract

The invention discloses a mobile robot control device giving consideration to field navigation patrol and transportation load, and relates to the field of robot control, and the device comprises a sensor data collection module which forms multi-source sensor data; the environment understanding module is used for predicting terrain risks and weather influences based on the feature vectors and outputting environment feature data and risk prediction; the navigation control module is used for navigating a path based on the environment characteristic data and generating a robot motion control instruction; the load management module adjusts parameters of the navigation control module based on the load state; the feedback adjusting module is used for dynamically adjusting control parameters according to the load state and feeding back the control parameters to the navigation control module to optimize and modify a robot motion control instruction; and the intelligent detection module is used for detecting environment abnormity and dangerous case events and outputting an alarm signal. The method has the advantages that accurate prediction of environmental risks is realized through multi-sensor layered feature extraction and attention weighted fusion, and motion parameters are dynamically adjusted in combination with a moment balance model.

Description

Technical Field

[0001] The present invention relates to the field of robot control, and in particular to a mobile robot control device capable of navigating and inspecting in the field and transporting loads. Background Art

[0002] With the widespread application of mobile robot technology in fields such as field exploration, material transportation, and environmental inspections, autonomous navigation and payload transport capabilities have become critical requirements. However, existing mobile robot control systems fail in complex field environments due to the ineffectiveness of their visual sensors in rainy and foggy conditions, and their lidar systems fail to accurately predict the collapse risk of soft surfaces (such as muddy swamps). This results in path planning ignoring environmental risk factors (sudden slope changes, collapse probability), increasing operational risks. Furthermore, during movement, motion parameters (speed, curvature radius) cannot be adjusted in real time, which can easily cause the robot to capsize, especially on steep slopes or in strong winds. Summary of the Invention

[0003] In order to solve the above technical problems, a mobile robot control device is provided that takes into account both field navigation and inspection and transport loads. This technical solution solves the problems raised in the above background technology, such as one-sided environmental perception, poor adaptability to dynamic changes in loads, inefficient multi-source data fusion, and insufficient real-time response in complex field environments.

[0004] In order to achieve the above objects, the technical solution adopted by the present invention is: A mobile robot control device capable of navigating and inspecting in the field and transporting a load, comprising: The sensor data acquisition module acquires visual data, lidar data, IMU data and environmental sensor data in the field environment through visual sensors, lidar, IMU data and environmental sensor data to form multi-source sensor data; The environmental understanding module is connected to the sensor data acquisition module, receives and processes multi-source sensor data, fuses multi-sensor data through a deep learning model, extracts features and recognizes patterns, and predicts terrain risks and weather impacts based on feature vectors, outputting environmental feature data and risk predictions; The navigation control module is connected to the environment understanding module, and it performs navigation path based on the environmental feature data and generates robot motion control instructions; The load management module monitors the transport load status in real time and adjusts the parameters of the navigation control module based on the load status; The feedback adjustment module is connected to the navigation control module and the load management module, dynamically adjusts the control parameters according to the load status, and feeds back to the navigation control module to optimize and modify the robot motion control instructions; The intelligent detection module is connected to the sensor data acquisition module and the environment understanding module to detect environmental anomalies and dangerous events and output alarm signals.

[0005] Preferably, the method of fusing multi-sensor data through a deep learning model, extracting features and performing pattern recognition, and predicting terrain risks and weather impacts based on feature vectors to output environmental feature data and risk predictions specifically includes: Synchronize visual data, lidar data, IMU data, and environmental sensor data in time and space, and unify heterogeneous data into the robot's coordinate system through a coordinate transformation matrix; Perform hierarchical feature extraction on the internal data of the robot body coordinate system to obtain feature data; Based on the dynamic weighted fusion step of the attention mechanism, the multimodal feature vector output is used to generate the environment joint feature vector; The environmental joint feature vector is input into the multi-layer perceptron to output the terrain risk level and weather impact coefficient. The terrain risk level includes slope mutation and collapse probability, and the weather impact coefficient includes rain and fog visibility attenuation and wind speed disturbance intensity.

[0006] Preferably, the step of performing hierarchical feature extraction on the internal data of the robot body coordinate system to obtain feature data specifically includes: The multi-source sensor data after spatiotemporal synchronization is classified and processed by sensor type to obtain visual data, lidar data, IMU data and environmental sensor data; Use convolutional neural networks to extract spatial texture features from visual data; Extract three-dimensional structural features from lidar data through point cloud network; Extract motion state features from IMU data through time series model; Extract statistical features from environmental sensor data; The above extracted statistical features are hierarchically encoded according to sensor types to generate multimodal feature vectors.

[0007] Preferably, the step of dynamically weighted fusion based on the attention mechanism outputs a multimodal feature vector to generate an environment joint feature vector, specifically comprising: The three-dimensional structural feature vector extracted from the lidar point cloud is used as the benchmark reference vector; Calculate the cosine similarity between the spatial texture feature vector extracted by the visual sensor, the motion state feature vector extracted by the IMU sensor, and the statistical feature vector extracted by the environmental sensor and the reference vector, and generate the attention weight coefficient of each modal feature; The multimodal feature vectors are dynamically weighted and fused according to the attention weight coefficient to generate the environment joint feature vector. The attention weight coefficient is determined by calculating the cosine similarity between the spatial texture feature vector extracted by the visual sensor, the motion state feature vector extracted by the IMU sensor, and the statistical feature vector extracted by the environmental sensor and the three-dimensional structure feature vector extracted by the lidar point cloud; The fused environmental joint feature vector is output to the risk prediction module for the prediction of terrain risk level and weather impact coefficient.

[0008] Preferably, the navigation path based on the environmental feature data and the generation of robot motion control instructions specifically include: Receive environmental characteristic data, terrain risk level, and weather impact coefficient output by the environmental understanding module; Construct a cost function that integrates environmental risks and motion constraints, with the robot's current position as the starting point, the target point as the end point, and the grid map as the search space; When the slope mutation risk quantification value is greater than the maximum slope tolerance or the wind speed disturbance cost is greater than the collapse probability safety threshold, the node is directly eliminated and the initial navigation path is generated. The maximum slope tolerance and collapse probability safety threshold are obtained through historical experimental data statistics. The initial path is locally optimized in real time using the dynamic window method to generate motion control instructions including linear velocity and angular velocity; The motion control instructions are output to the robot chassis actuator, and the load status parameters fed back by the load management module are simultaneously received to dynamically adjust the path curvature radius threshold and maximum speed.

[0009] Preferably, the cost function for constructing the fusion of environmental risk and motion constraint specifically includes: ; in, is the output value of the cost function that integrates environmental risk and motion constraints, is the Euclidean distance cost between nodes, Quantify the risk of sudden slope changes in medium terrain risk level, is the collapse probability value in the medium terrain risk level, is the visibility attenuation cost due to rain and fog, is the wind speed disturbance cost, α, β, γ, δ are the attention weight coefficients.

[0010] Preferably, the real-time monitoring of the transport load status and adjusting the parameters of the navigation control module based on the load status specifically includes: The load distribution state is detected in real time through the pressure sensor array, and the load center of gravity offset is calculated; When the center of gravity offset exceeds a preset safety threshold, a deceleration command is sent to the navigation control module, and the path curvature radius threshold is dynamically reduced. The preset safety threshold is determined by statistically analyzing the robot's tipping over threshold under different load distributions using historical experimental data, combined with the current terrain risk level and weather impact coefficient. Based on the load weight distribution data, the robot chassis attitude angle is adjusted through the torque balance model, and the chassis tilt compensation instruction is generated and synchronized with the motion control instruction.

[0011] Preferably, the dynamically adjusting control parameters according to the load state specifically includes: Obtain center of gravity offset and load weight distribution data from the load management module; Calculate the speed reduction proportional factor based on the proportional relationship between the center of gravity offset and the preset safety threshold; Based on the size of the center of gravity offset, the curvature radius reduction proportional factor is calculated; Based on the load weight distribution data, the robot chassis tilt compensation angle is calculated through the torque balance model; The speed reduction proportional factor, curvature radius reduction proportional factor and tilt compensation angle are used as control parameter adjustment quantities and fed back to the navigation control module to adjust the linear speed, angular velocity and chassis posture of the robot motion control instructions in real time.

[0012] Compared with the prior art, the present invention has the following beneficial effects: The present invention proposes a mobile robot control device that takes into account both field navigation and inspection and load transportation. Using the three-dimensional structural features of a laser radar as a reference vector, the device calculates the similarity between other modal features and their cosines to generate dynamic weights. This weighted fusion generates a joint environmental feature vector, which is then input into a multi-layer perceptron (MLP) to output terrain risk levels (sudden slope change, collapse probability) and weather impact coefficients (visibility attenuation due to rain and fog, wind speed disturbance). This automatically reduces the visual weight in severe weather, enhancing the contributions of laser and environmental sensors. This reduces terrain risk prediction errors and improves collapse probability identification accuracy in rainy and foggy scenarios. Furthermore, layered processing and an attention mechanism shorten feature extraction time, meeting real-time response requirements. The dynamic weighting mechanism based on laser features can also effectively address sensor data conflicts. This invention proposes a mobile robot control device that combines field navigation, inspection, and payload transportation. This device uses a pressure sensor array to monitor load distribution in real time, calculate center of gravity offset and weight distribution, and preset safety thresholds that integrate historical tipping thresholds, real-time terrain risk levels (sudden slope changes), and weather impact factors (wind speed fluctuations). When the center of gravity offset exceeds the threshold, a speed reduction command is triggered and the path curvature radius threshold is reduced. Simultaneously, based on the load weight distribution data, the chassis tilt compensation angle is calculated using a rigid body dynamics model. Multi-parameter coordinated adjustment is performed. When the slope risk exceeds level 2 or higher, the safety threshold is automatically tightened, dynamically adapting to complex terrain. The speed is linearly reduced to ensure stability, and the curvature radius is exponentially reduced to limit centrifugal forces (for example, when the offset reaches 120% of the threshold, the curvature radius is reduced to 80%). Furthermore, the tilt angle θ is calculated using a torque balance model (e.g., θ = 0.1 rad), actively adjusting the chassis posture. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 It is a schematic diagram of the process of the present invention; Figure 2 This is a flowchart of the present invention that uses a deep learning model to fuse multi-sensor data, extract features and perform pattern recognition, and predict terrain risks and weather impacts based on feature vectors to output environmental feature data and risk predictions; Figure 3 This is a schematic diagram of the process of performing hierarchical feature extraction on the internal data of the robot body coordinate system to obtain feature data in the present invention; Figure 4 This is a flow chart of generating an environment joint feature vector from the multimodal feature vectors outputted by the dynamic weighted fusion step of the attention mechanism in the present invention; Figure 5 A schematic diagram of the process of navigating the path based on environmental feature data and generating robot motion control instructions in the present invention; Figure 6 Schematic diagram of the process of real-time monitoring of the transport load status and adjusting the parameters of the navigation control module based on the load status in the present invention; Figure 7 It is a schematic diagram of the process of dynamically adjusting the control parameters according to the load status in the present invention. DETAILED DESCRIPTION

[0014] The following description is intended to disclose the present invention so that those skilled in the art can implement the present invention. The preferred embodiments described below are merely examples, and those skilled in the art may conceive of other obvious variations.

[0015] Reference Figure 1 As shown, a mobile robot control device that takes into account both field navigation and inspection and load transportation includes: The sensor data acquisition module acquires visual data, lidar data, IMU data and environmental sensor data in the field environment through visual sensors, lidar, IMU data and environmental sensor data to form multi-source sensor data; The environmental understanding module is connected to the sensor data acquisition module, receives and processes multi-source sensor data, fuses multi-sensor data through a deep learning model, extracts features and recognizes patterns, and predicts terrain risks and weather impacts based on feature vectors, outputting environmental feature data and risk predictions; The navigation control module is connected to the environment understanding module, and it performs navigation path based on the environmental feature data and generates robot motion control instructions; The load management module monitors the transport load status in real time and adjusts the parameters of the navigation control module based on the load status; The feedback adjustment module is connected to the navigation control module and the load management module, dynamically adjusts the control parameters according to the load status, and feeds back to the navigation control module to optimize and modify the robot motion control instructions; The intelligent detection module is connected to the sensor data acquisition module and the environment understanding module to detect environmental anomalies and dangerous events and output alarm signals.

[0016] The sensor data acquisition module integrates visual sensors (such as RGB-D cameras), lidar (such as the Velodyne series), inertial measurement unit (IMU) (such as the MPU9250), and environmental sensors (temperature, humidity, and barometer) through hardware, enabling the simultaneous collection of multi-source heterogeneous data. The environmental understanding module employs a hierarchical feature extraction architecture: First, a coordinate transformation matrix is ​​used to align the heterogeneous data spatially and temporally to the robot's coordinate system. Next, a dedicated network processes the sensor data separately: ResNet is used to extract spatial texture features from visual data, PointNet++ is used to extract 3D structural features from lidar data, LSTM is used to extract motion state features from IMU data, and statistical features from environmental sensors. Multimodal feature vectors are then dynamically weighted and fused using an attention mechanism (using lidar features as a baseline and calculating the cosine similarity between other modal features and the vector as weights) to generate a joint environmental feature vector. Finally, a multi-layer perceptron is used to output the terrain risk level (including a quantified slope change and a collapse probability) and the weather impact factor (including visibility loss due to rain and fog, and wind speed disturbance intensity). The navigation control module constructs a cost function that incorporates environmental risks, directly eliminating path nodes to generate an initial navigation path. This path is then optimized in real time using a dynamic windowing method, outputting motion control commands that include linear and angular velocities. The load management module uses a pressure sensor array to detect load distribution and calculate the center of gravity offset. When the offset exceeds a preset safety threshold (dynamically adjusted based on terrain risk and weather factors), it triggers a speed reduction command and reduces the path curvature radius threshold. Simultaneously, it calculates the robot chassis tilt compensation angle based on a torque balance model. The feedback control module calculates the speed reduction factor, curvature radius reduction factor, and tilt compensation angle based on the proportional relationship between the center of gravity offset and the safety threshold, adjusting the navigation control module's motion command parameters in real time. The intelligent detection module uses an anomaly pattern recognition algorithm (such as the isolation forest algorithm) to jointly analyze the raw sensor data and the risk prediction output by the environmental understanding module, triggering an environmental hazard alert.

[0017] Therefore, during its use, dynamic weighted fusion (cosine similarity weight distribution) can be achieved through the attention mechanism, which reduces the error in environmental feature extraction and improves the accuracy of terrain risk prediction. At the same time, the dynamic adjustment mechanism of the environmental risk weight coefficient in the cost function improves the robot's path avoidance success rate when the slope suddenly increases to more than 15° or the wind speed increases to more than 8m / s, making the navigation solution more adaptable to complex terrain in the wild; the chassis posture compensation algorithm based on the torque balance model improves the load offset tolerance and reduces the number of emergency stops in the center of gravity offset processing solution.

[0018] Reference Figure 2 As shown, the method of fusing multi-sensor data through deep learning models, extracting features and recognizing patterns, and predicting terrain risks and weather impacts based on feature vectors, outputting environmental feature data and risk predictions specifically includes: Synchronize visual data, lidar data, IMU data, and environmental sensor data in time and space, and unify heterogeneous data into the robot's coordinate system through a coordinate transformation matrix; Perform hierarchical feature extraction on the internal data of the robot body coordinate system to obtain feature data; Based on the dynamic weighted fusion step of the attention mechanism, the multimodal feature vector output is used to generate the environment joint feature vector; The environmental joint feature vector is input into the multi-layer perceptron to output the terrain risk level and weather impact coefficient. The terrain risk level includes slope mutation and collapse probability, and the weather impact coefficient includes rain and fog visibility attenuation and wind speed disturbance intensity.

[0019] A hardware timestamp alignment mechanism (such as the PTP protocol) is used to achieve synchronous acquisition of multi-source sensor data. The visual data (RGB-D camera), lidar point cloud, IMU pose and environmental sensor (temperature, humidity, barometer) data are uniformly converted to the robot's body coordinate system through a coordinate transformation matrix (homogeneous transformation matrix). Among them: the lidar point cloud is converted to the body coordinate system through an external parameter calibration matrix; the IMU data is aligned with the body coordinate system after pre-integration compensation for motion distortion; the visual data is projected into the same coordinate system in combination with the camera's intrinsic parameters and depth information. That is, visual data can: use convolutional neural networks (such as ResNet) to extract spatial texture features to capture environmental details such as surface vegetation and rock distribution; lidar data can extract three-dimensional structural features through point cloud networks (such as PointNet++) to quantify terrain undulations and obstacle contours; IMU data can use time series models (such as LSTM) to extract motion state features (acceleration, angular velocity changes); environmental sensor data can characterize dynamic changes in temperature, humidity and air pressure by calculating statistical features (mean, variance); feature encoding allows the above features to be hierarchically encoded into multimodal feature vectors according to sensor type (for example, visual feature dimension 512, laser feature dimension 256), complete hierarchical feature extraction, retain modal specificity (vision focuses on texture, laser focuses on structure), and reduce fusion error through adaptive weighting of the attention mechanism (such as dynamically assigning weights through cosine similarity). At the same time, hierarchical feature extraction compresses redundant information, reduces the MLP input dimension, and shortens the inference time.

[0020] Then, using the LiDAR's 3D structural feature vector as a baseline reference vector, the cosine similarity between the visual texture features, IMU motion features, and environmental statistical features is calculated, generating attention weight coefficients for each modality. Based on these weight coefficients, the multimodal features are dynamically weighted and fused to generate a joint environmental feature vector. This joint environmental feature vector is then input into a multi-layer perceptron (MLP) to output a terrain risk level. The prediction layer uses a Sigmoid activation function to normalize the risk value to the [0, 1] range. The similarity-based attention mechanism automatically reduces the visual weight in rainy and foggy weather, increasing the weight of the laser and environmental sensors, thereby improving the accuracy of terrain risk prediction.

[0021] Reference Figure 3 As shown, the hierarchical feature extraction of the internal data of the robot body coordinate system is performed to obtain the feature data specifically including: The multi-source sensor data after spatiotemporal synchronization is classified and processed by sensor type to obtain visual data, lidar data, IMU data and environmental sensor data; Use convolutional neural networks to extract spatial texture features from visual data; Extract three-dimensional structural features from lidar data through point cloud network; Extract motion state features from IMU data through time series model; Extract statistical features from environmental sensor data; The above extracted statistical features are hierarchically encoded according to sensor types to generate multimodal feature vectors.

[0022] After completing the spatiotemporal synchronization of multi-source sensor data (through hardware timestamp alignment and coordinate transformation matrix unification to the robot's coordinate system), the data is first classified by sensor type into visual data, lidar data, IMU data, and environmental sensor data. For visual data (e.g., RGB-D images), convolutional neural networks (e.g., ResNet) are used to extract spatial texture features to capture environmental details such as surface vegetation and rock distribution. For lidar point cloud data, point cloud networks (e.g., PointNet++) are used to extract 3D structural features to quantify terrain undulations and obstacle contours. For IMU data (acceleration and angular velocity), time series models (e.g., LSTM) are used to extract motion state features to characterize changes in the robot's motion. For environmental sensor data (e.g., temperature, humidity, and air pressure), statistical features (mean and variance) are directly calculated to reflect environmental dynamics. Finally, the extracted visual texture features (512 dimensions), lidar structural features (256 dimensions), IMU motion features, and environmental statistical features are hierarchically encoded by sensor type to generate a unified multimodal feature vector.

[0023] During its construction, dedicated networks are selected strictly according to the physical characteristics of the sensor (CNN extracts spatial textures, point cloud networks process three-dimensional structures, and LSTM captures temporal patterns), retaining the specificity of each modality (such as vision focusing on textures and laser focusing on geometric structures) to avoid feature confusion. During its implementation, redundant information is compressed before hierarchical encoding (such as CNN and point cloud networks automatically filter out invalid backgrounds), significantly reducing feature dimensions (compared to original point cloud / image data), reducing the input dimensions of subsequent multimodal fusion, and improving inference speed.

[0024] Then, through a highly targeted dedicated network, the CNN's recognition error of visual texture can be reduced, and the point cloud network's quantification accuracy of terrain undulations can reach the centimeter level; at the same time, layered parallel processing (visual / laser / IMU features can be extracted synchronously) is less time-consuming than the serial fusion method; and layered encoding retains the original sensor correlation, providing structured input for subsequent multimodal fusion based on the attention mechanism (such as weighting based on laser features), thereby improving the accuracy of terrain risk prediction.

[0025] Reference Figure 4 As shown, the dynamic weighted fusion step based on the attention mechanism outputs a multimodal feature vector and generates an environmental joint feature vector specifically including: The three-dimensional structural feature vector extracted from the lidar point cloud is used as the benchmark reference vector; Calculate the cosine similarity between the spatial texture feature vector extracted by the visual sensor, the motion state feature vector extracted by the IMU sensor, and the statistical feature vector extracted by the environmental sensor and the reference vector, and generate the attention weight coefficient of each modal feature; The multimodal feature vectors are dynamically weighted and fused according to the attention weight coefficient to generate the environment joint feature vector. The attention weight coefficient is determined by calculating the cosine similarity between the spatial texture feature vector extracted by the visual sensor, the motion state feature vector extracted by the IMU sensor, and the statistical feature vector extracted by the environmental sensor and the three-dimensional structure feature vector extracted by the lidar point cloud; The fused environmental joint feature vector is output to the risk prediction module for the prediction of terrain risk level and weather impact coefficient.

[0026] The three-dimensional structural feature vector extracted from the lidar point cloud data through the point cloud network (such as PointNet++) is used as the baseline reference vector. After the lidar point cloud is unified to the robot body coordinate system through the coordinate transformation matrix, the 256-dimensional feature vector is extracted hierarchically through the point cloud network as the benchmark. The baseline reference vector represents the geometric structure of the environment (such as terrain undulations and obstacle contours). Because it is less affected by light and weather and has better stability than other modal features, the combined feature vector more comprehensively represents environmental risks (such as the probability of terrain collapse and wind speed disturbances).

[0027] After spatiotemporal synchronization, multi-source data is processed by sensor type. After each modal feature is independently encoded, the cosine similarity between the visual sensor (512-dimensional spatial texture features extracted by CNN), the IMU (motion state features extracted by LSTM), and the environmental sensor (statistical features such as the mean and variance of temperature and humidity) and the baseline vector is calculated in real time. The multimodal feature vectors are weighted and summed according to the weight coefficient to generate a joint environmental feature vector, which is then output to the risk prediction module. Dynamic weighting based on cosine similarity allows the system to automatically reduce the visual weight in inclement weather, enhance the contributions of the laser and environmental sensors, and reduce feature fusion errors.

[0028] Reference Figure 5 As shown, the navigation path based on the environmental feature data and the generation of robot motion control instructions specifically include: Receive environmental characteristic data, terrain risk level, and weather impact coefficient output by the environmental understanding module; Construct a cost function that integrates environmental risks and motion constraints, with the robot's current position as the starting point, the target point as the end point, and the grid map as the search space; When the slope mutation risk quantification value is greater than the maximum slope tolerance or the wind speed disturbance cost is greater than the collapse probability safety threshold, the node is directly eliminated and the initial navigation path is generated. The maximum slope tolerance and collapse probability safety threshold are obtained through historical experimental data statistics. The initial path is locally optimized in real time using the dynamic window method to generate motion control instructions including linear velocity and angular velocity; The motion control instructions are output to the robot chassis actuator, and the load status parameters fed back by the load management module are simultaneously received to dynamically adjust the path curvature radius threshold and maximum speed.

[0029] The flight control module receives real-time environmental feature data (including terrain texture, obstacle distribution, etc.), terrain risk level, and weather impact coefficients transmitted by the environment understanding module via a hardware bus (such as CAN or EtherCAT). This data then constructs a cost function that integrates environmental risk with motion constraints. Before the A* algorithm traverses the grid map, thresholds (such as maximum slope tolerance and collapse probability safety threshold) are set based on historical experimental statistics. If a node meets the predicate, it is directly eliminated, generating an initial path free of high-risk nodes. Using this initial path as a reference, feasible trajectories are sampled in velocity space (linear velocity v, angular velocity ω). By evaluating the trajectory's obstacle distance, path conformity, and dynamic feasibility, smooth motion control commands are generated. When these motion commands are output to the chassis actuators (such as the motor drive), they simultaneously receive pressure sensor data from the load management module for load status feedback.

[0030] During its use, by pre-screening high-risk nodes (such as areas with sudden slope changes >15° or wind speeds >8m / s), path planning time can be reduced and the avoidance success rate can be improved; at the same time, the attention weight coefficients α, β, γ, and δ are dynamically adjusted according to the confidence of multiple sensors (such as automatically increasing the γ weight in rainy and foggy days), so that the cost function can adapt to environmental changes in real time, and the positioning accuracy is improved compared with fixed weight models (such as pure distance optimization); and the load state feedback mechanism calculates the robot chassis tilt compensation angle (such as θ=0.1rad) in real time through the torque balance model, and synchronously adjusts the motion command parameters to reduce the number of emergency stops during heavy-load transportation, and the path tracking error is controlled within ±0.1m (traditional methods easily lead to overturning due to ignoring load offset).

[0031] The cost function for constructing the fusion of environmental risk and motion constraints specifically includes: ; in, is the output value of the cost function that integrates environmental risk and motion constraints, is the Euclidean distance cost between nodes, Quantify the risk of sudden slope changes in medium terrain risk level, is the collapse probability value in the medium terrain risk level, is the visibility attenuation cost due to rain and fog, is the wind speed disturbance cost, α, β, γ, δ are the attention weight coefficients.

[0032] Reference Figure 6 As shown, the real-time monitoring of the transport load status and adjusting the parameters of the navigation control module based on the load status specifically include: The load distribution state is detected in real time through the pressure sensor array, and the load center of gravity offset is calculated; When the center of gravity offset exceeds a preset safety threshold, a deceleration command is sent to the navigation control module, and the path curvature radius threshold is dynamically reduced. The preset safety threshold is determined by statistically analyzing the robot's tipping over threshold under different load distributions using historical experimental data, combined with the current terrain risk level and weather impact coefficient. Based on the load weight distribution data, the robot chassis attitude angle is adjusted through the torque balance model, and the chassis tilt compensation instruction is generated and synchronized with the motion control instruction.

[0033] A high-density array of pressure sensors (such as piezoresistive or piezoelectric sensors) is integrated into the bottom of the robot transport platform, forming a grid-like distribution. Each sensor node collects local pressure values ​​in real time and transmits them to the embedded processor of the load management module via a bus (such as CAN or SPI). Preset safety thresholds are established by statistically analyzing historical experimental data to create a database that stores the robot's tipping thresholds under different load distributions (e.g., eccentricity and eccentricity). This module dynamically calculates the safety threshold for the current operating conditions by combining the terrain risk level (e.g., slope mutation level, collapse probability) and weather influence factors (e.g., wind speed disturbance intensity) output by the environmental understanding module in real time. For example, when the slope mutation risk level is ≥ Level 2 and the wind speed disturbance factor is > 0.5, the safety threshold is tightened to 80% of the historical threshold. Using a built-in torque balance model based on rigid-body dynamics, the load weight distribution data (generated from the raw sensor array data after Kalman filtering and denoising) is input to calculate the torque distribution of each drive wheel of the robot chassis and output a chassis tilt compensation angle command.

[0034] Therefore, during use, the threshold can be dynamically adjusted by integrating terrain risk and weather coefficients to reduce the risk of the robot overturning; at the same time, the overturning moment is offset by active tilt compensation of the chassis, reducing the number of emergency stops; and the load management module forms a closed loop with the navigation control module and feedback adjustment module.

[0035] Reference Figure 7 As shown, the dynamic adjustment of control parameters according to the load state specifically includes: Obtain center of gravity offset and load weight distribution data from the load management module; Calculate the speed reduction proportional factor based on the proportional relationship between the center of gravity offset and the preset safety threshold; Based on the size of the center of gravity offset, the curvature radius reduction proportional factor is calculated; Based on the load weight distribution data, the robot chassis tilt compensation angle is calculated through the torque balance model; The speed reduction proportional factor, curvature radius reduction proportional factor and tilt compensation angle are used as control parameter adjustment quantities and fed back to the navigation control module to adjust the linear speed, angular velocity and chassis posture of the robot motion control instructions in real time.

[0036] The preset safety threshold is calculated through historical experiments to determine the critical value of the robot's overturning under different load distributions. It is dynamically adjusted by integrating the terrain risk level (slope mutation level) and weather impact coefficient (wind speed disturbance intensity) output in real time by the environmental understanding module (for example, the threshold is tightened by 20% when the slope risk is ≥ level 2). When the center of gravity offset exceeds the threshold, the speed reduction factor increases linearly according to the proportion of the offset exceeding the standard (for example, when the offset is 120% of the threshold, the speed is reduced to 80%). The curvature radius reduction factor decreases exponentially with the increase in the offset, limiting the centrifugal force of the robot when driving on a curve. The above parameters (speed factor, curvature factor, and tilt angle) are input into the motion command optimizer of the navigation control module in real time to dynamically adjust the linear speed, angular velocity, and chassis servo angle to form a "perception-computation-execution" closed loop.

[0037] The speed, steering curvature and chassis posture are optimized simultaneously, and the overturning torque is actively offset by the torque balance model, which reduces the number of emergency stops for heavy-load transportation. The preset safety threshold integrates real-time environmental risks (such as collapse probability and wind speed disturbance), which improves stability in complex terrain compared to the fixed threshold method (such as relying solely on historical data). The path tracking error is controlled within ±0.1m, and the proportional factor is directly generated through linear / nonlinear functions to avoid complex optimization algorithms (such as model predictive control MPC). The processing delay on the embedded platform is less than 10ms, meeting the real-time response requirements in the field. At the same time, the tight coupling mechanism of load state and navigation control (such as the dynamic reduction of the curvature radius threshold) improves the robot's center of gravity offset tolerance and significantly reduces the risk of rollover on sloped terrain.

[0038] In summary, the advantages of the present invention are: accurate prediction of environmental risks is achieved through multi-sensor hierarchical feature extraction and attention weighted fusion, and motion parameters are dynamically adjusted in combination with the torque balance model to improve navigation robustness, load offset tolerance and real-time response capabilities in complex terrain.

[0039] The above shows and describes the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The above embodiments and descriptions merely illustrate the principles of the present invention. Various changes and modifications may be made to the present invention without departing from the spirit and scope of the present invention. Such changes and modifications are intended to fall within the scope of the present invention. The scope of protection claimed by the present invention is defined by the appended claims and their equivalents.

Claims

1. A mobile robot control device that takes into account both field navigation and inspection and transport loads, characterized in that: include: The sensor data acquisition module acquires visual data, lidar data, IMU data and environmental sensor data in the field environment through visual sensors, lidar, IMU data and environmental sensor data to form multi-source sensor data; The environmental understanding module is connected to the sensor data acquisition module, receives and processes multi-source sensor data, fuses multi-sensor data through a deep learning model, extracts features and recognizes patterns, and predicts terrain risks and weather impacts based on feature vectors, outputting environmental feature data and risk predictions; The navigation control module is connected to the environment understanding module, and it performs navigation path based on the environmental feature data and generates robot motion control instructions; The load management module monitors the transport load status in real time and adjusts the parameters of the navigation control module based on the load status; The feedback adjustment module is connected to the navigation control module and the load management module, dynamically adjusts the control parameters according to the load status, and feeds back to the navigation control module to optimize and modify the robot motion control instructions; The intelligent detection module is connected to the sensor data acquisition module and the environment understanding module to detect environmental anomalies and dangerous events and output alarm signals.

2. A mobile robot control device capable of navigating and inspecting in the field and transporting loads according to claim 1, characterized in that: The method of fusing multi-sensor data through a deep learning model, extracting features and performing pattern recognition, and predicting terrain risks and weather impacts based on feature vectors to output environmental feature data and risk predictions specifically includes: Synchronize visual data, lidar data, IMU data, and environmental sensor data in time and space, and unify heterogeneous data into the robot's coordinate system through a coordinate transformation matrix; Perform hierarchical feature extraction on the internal data of the robot body coordinate system to obtain feature data; Based on the dynamic weighted fusion step of the attention mechanism, the multimodal feature vector output is used to generate the environment joint feature vector; The environmental joint feature vector is input into the multi-layer perceptron to output the terrain risk level and weather impact coefficient. The terrain risk level includes slope mutation and collapse probability, and the weather impact coefficient includes rain and fog visibility attenuation and wind speed disturbance intensity.

3. The mobile robot control device for field navigation and inspection and transporting loads according to claim 2, characterized in that: The hierarchical feature extraction of the internal data of the robot body coordinate system to obtain the feature data specifically includes: The multi-source sensor data after spatiotemporal synchronization is classified and processed by sensor type to obtain visual data, lidar data, IMU data and environmental sensor data; Use convolutional neural networks to extract spatial texture features from visual data; Extract three-dimensional structural features from lidar data through point cloud network; Extract motion state features from IMU data through time series model; Extract statistical features from environmental sensor data; The above extracted statistical features are hierarchically encoded according to sensor types to generate multimodal feature vectors.

4. The mobile robot control device for field navigation and inspection and transporting loads according to claim 3, characterized in that: The multimodal feature vector outputted by the dynamic weighted fusion step based on the attention mechanism and the generation of the environment joint feature vector specifically include: The three-dimensional structural feature vector extracted from the lidar point cloud is used as the benchmark reference vector; Calculate the cosine similarity between the spatial texture feature vector extracted by the visual sensor, the motion state feature vector extracted by the IMU sensor, and the statistical feature vector extracted by the environmental sensor and the reference vector, and generate the attention weight coefficient of each modal feature; The multimodal feature vectors are dynamically weighted and fused according to the attention weight coefficient to generate the environment joint feature vector. The attention weight coefficient is determined by calculating the cosine similarity between the spatial texture feature vector extracted by the visual sensor, the motion state feature vector extracted by the IMU sensor, and the statistical feature vector extracted by the environmental sensor and the three-dimensional structure feature vector extracted by the lidar point cloud; The fused environmental joint feature vector is output to the risk prediction module for the prediction of terrain risk level and weather impact coefficient.

5. The mobile robot control device for field navigation and inspection and transporting loads according to claim 4, characterized in that: The navigation path based on the environmental feature data and the generation of robot motion control instructions specifically include: Receive environmental characteristic data, terrain risk level, and weather impact coefficient output by the environmental understanding module; Construct a cost function that integrates environmental risks and motion constraints, with the robot's current position as the starting point, the target point as the end point, and the grid map as the search space; When the slope mutation risk quantification value is greater than the maximum slope tolerance or the wind speed disturbance cost is greater than the collapse probability safety threshold, the node is directly eliminated and the initial navigation path is generated. The maximum slope tolerance and collapse probability safety threshold are obtained through historical experimental data statistics. The initial path is locally optimized in real time using the dynamic window method to generate motion control instructions including linear velocity and angular velocity; The motion control instructions are output to the robot chassis actuator, and the load status parameters fed back by the load management module are simultaneously received to dynamically adjust the path curvature radius threshold and maximum speed.

6. The mobile robot control device for field navigation and inspection and transporting loads according to claim 5, characterized in that: The cost function for constructing the fusion of environmental risk and motion constraints specifically includes: ； in, is the output value of the cost function that integrates environmental risk and motion constraints, is the Euclidean distance cost between nodes, Quantify the risk of sudden slope changes in medium terrain risk level, is the collapse probability value in the medium terrain risk level, is the visibility attenuation cost due to rain and fog, is the wind speed disturbance cost, α, β, γ, δ are the attention weight coefficients.

7. The mobile robot control device for field navigation and inspection and transporting loads according to claim 6, characterized in that: The real-time monitoring of the transport load status and adjusting the parameters of the navigation control module based on the load status specifically includes: The load distribution state is detected in real time through the pressure sensor array, and the load center of gravity offset is calculated; When the center of gravity offset exceeds a preset safety threshold, a deceleration command is sent to the navigation control module, and the path curvature radius threshold is dynamically reduced. The preset safety threshold is determined by statistically analyzing the robot's tipping over threshold under different load distributions using historical experimental data, combined with the current terrain risk level and weather impact coefficient. Based on the load weight distribution data, the robot chassis attitude angle is adjusted through the torque balance model, and the chassis tilt compensation instruction is generated and synchronized with the motion control instruction.

8. The mobile robot control device capable of both field navigation and inspection and transporting loads according to claim 7, characterized in that: The dynamic adjustment of control parameters according to the load state specifically includes: Obtain center of gravity offset and load weight distribution data from the load management module; Calculate the speed reduction proportional factor based on the proportional relationship between the center of gravity offset and the preset safety threshold; Based on the size of the center of gravity offset, the curvature radius reduction proportional factor is calculated; Based on the load weight distribution data, the robot chassis tilt compensation angle is calculated through the torque balance model; The speed reduction proportional factor, curvature radius reduction proportional factor and tilt compensation angle are used as control parameter adjustment quantities and fed back to the navigation control module to adjust the linear speed, angular velocity and chassis posture of the robot motion control instructions in real time.

Images