AI large model driven explosive ordnance disposal robot grabbing method and system

Abstract

The invention discloses an AI large model driven explosive ordnance disposal robot grabbing method and system, and relates to the technical field of explosive ordnance disposal robots, the system comprises an intelligent identification module, the intelligent identification module is used for carrying out real-time analysis on environment data collected by a robot through an AI large model, and the intelligent identification module is used for identifying the shape, material and position characteristics of explosives; and the self-adaptive grabbing strategy module is used for dynamically generating a grabbing path, a grabbing angle and grabbing force parameters according to the recognition result. According to the invention, through deep integration of multi-modal perception, intelligent decision and adaptive control, significant beneficial effects are brought, the system uses the visual model based on Transform, even under the condition that an explosive is severely shielded, pixel-level accurate identification and three-dimensional reconstruction can be realized, and the system has high practicability. High-precision geometric and physical information with uncertainty evaluation is provided for grabbing.

Description

Technical Field

[0001] This application relates to the technical field of bomb disposal robots, and in particular to a method and system for grasping bomb disposal robots driven by AI large-scale models. Background Technology

[0002] The core task of bomb disposal robots is to safely and reliably transfer or dismantle suspected explosive devices while minimizing secondary damage. The grasping operation is a crucial step in this process, and its success or failure directly affects the success or failure of the mission and on-site safety. However, existing bomb disposal robot grasping systems generally face the following technical bottlenecks: Traditional machine vision algorithms rely heavily on preset templates and ideal lighting conditions for identifying explosives in complex and unstructured bomb disposal environments. For explosives that are partially obscured, irregularly shaped, or made of special materials, the success rate of identification drops significantly, and they cannot provide accurate three-dimensional geometric and physical information that can be used for grasping planning. Most systems use pre-programmed grabbing paths and fixed force values, which lack adaptability and cannot dynamically calculate the optimal grabbing point and grabbing force according to the specific shape, material and posture of the explosive. This can easily lead to unstable grabbing, slippage of the object, or damage to the fragile structure of the explosive due to excessive force, resulting in catastrophic consequences. Faced with a dynamically changing environment and uncertainties in the process of the robot hand contacting objects, traditional control systems are slow to respond and struggle to achieve a compliant, adaptive force control closed loop, thus failing to guarantee the stability and safety of the grasping action. Existing systems typically place operators in a purely remote operation mode or a simple "execute / stop" mode, lacking transparency in the AI ​​decision-making process. Operators cannot quickly understand the AI's planning intentions and potential risks, making it difficult to intervene effectively at critical moments. They fail to deeply integrate human experience and judgment with AI's computing power. Summary of the Invention

[0003] To address the aforementioned issues, this application provides an AI-driven bomb disposal robot grasping method and system.

[0004] The AI-driven large-scale model-based bomb disposal robot grasping method and system provided in this application adopts the following technical solution: The AI-driven bomb disposal robot grasping system includes: The intelligent recognition module is used to analyze the environmental data collected by the robot in real time through an AI model, and is used to identify the shape, material and location characteristics of explosives. An adaptive grasping strategy module is used to dynamically generate grasping path, grasping angle and grasping force parameters based on the recognition results. The adaptive grasping strategy module is communicatively connected to the intelligent recognition module. A real-time environment perception module is used to monitor the interaction between the robot hand and the target object and environmental changes in real time through multi-sensor fusion technology. The real-time environment perception module is communicatively connected to the adaptive grasping strategy module. The grasping control module is used to control the robot hand to perform grasping actions based on the parameters output by the adaptive grasping strategy module and the feedback signal from the real-time environment perception module. The grasping control module is communicatively connected to the real-time environment perception module. The simulation verification and human-machine collaboration module is used to perform digital twin simulation verification of the strategy before physical grasping and realize human-machine collaborative decision-making through augmented reality interface. The simulation verification and human-machine collaboration module is communicatively connected to the grasping control module.

[0005] As a preferred technical solution of this application, the intelligent recognition module includes a visual analysis unit, a semantic segmentation unit, and a three-dimensional reconstruction unit. The visual analysis unit is used to extract multi-scale features from the video stream captured by the camera. The semantic segmentation unit uses a pre-trained model based on the Transformer architecture to accurately locate the outline of the explosive at the pixel level and predictively complete the occluded parts. The three-dimensional reconstruction unit is used to generate a high-precision three-dimensional geometric model and centroid distribution of the target object based on point cloud data.

[0006] As a preferred technical solution of this application, the adaptive grasping strategy module includes a grasping point optimization unit, a force control parameter generation unit, and an obstacle avoidance planning unit. The grasping point optimization unit is used to calculate the optimal set of grasping contact points and evaluate the grasping stability metric based on the geometric characteristics, estimated centroid, and physical stability of the explosive. The force control parameter generation unit is used to dynamically generate an adaptive grasping force curve by combining material identification results and historical data from torque sensors and by combining table lookup with online learning. The obstacle avoidance planning unit generates a collision-free smooth grasping trajectory based on a real-time updated voxel map and by using a fusion planning of an improved A* algorithm and a fast random tree algorithm.

[0007] As a preferred technical solution of this application, the real-time environment perception module includes a multi-sensor fusion unit, a dynamic obstacle tracking unit, and a contact state monitoring unit. The multi-sensor fusion unit uses a Kalman filter and deep learning fusion algorithm to simultaneously process data from the vision, touch, acoustic, and inertial measurement units and construct a unified environmental state estimate. The dynamic obstacle tracking unit detects and predicts the trajectory of moving obstacles and calculates the probability of collision risk based on target detection and optical flow. The contact state monitoring unit uses a high-precision torque sensor array to provide real-time feedback on torque changes during the grasping process.

[0008] As a preferred technical solution of this application, the simulation verification and human-machine collaboration module includes a digital twin unit and an augmented reality interaction unit. The digital twin unit constructs a virtual scene synchronized with the real environment in real time based on a physics engine to perform Monte Carlo simulations on the grasping strategy to be executed millions of times, and evaluates the grasping success rate and risk in a probabilistic form. The augmented reality interaction unit is used to overlay and display the grasping points, planned paths and force control parameters recommended by the AI ​​system on the operator's control interface and highlight the risk areas in a visual manner. The operator has the final decision-making power and can modify or confirm the AI ​​plan.

[0009] As a preferred technical solution of this application, the AI-driven bomb disposal robot grasping method, based on the AI-driven bomb disposal robot grasping system, includes the following steps: Step 1: Use AI big data models to fuse and analyze multimodal sensor data at the bomb disposal site to identify multidimensional features such as the type, shape, material, location, and attitude of explosives. Step 2: Based on the identified multidimensional features and real-time environmental parameters, an adaptive grasping strategy containing grasping points, paths, angles, and force curves is generated through a policy network driven by a reinforcement learning algorithm. Step 3: Construct a dynamic voxel map using multi-sensor fusion data, track obstacles in real time, and use model predictive control algorithms to optimize the grabbing path planning. Step four: Through a force feedback closed loop based on impedance control, the posture and force of the gripping mechanism are adjusted in real time to ensure the smoothness, stability and safety of the gripping action. Step 5: Before physical execution, the capture strategy is verified by ultra-real-time simulation in the digital twin system, and the optimal strategy is submitted to the operator for confirmation and execution through the AR interface.

[0010] As a preferred technical solution of this application, in step one, a multi-task learning framework is first adopted to process feature extraction, instance segmentation and anomaly detection tasks in parallel. Then, by introducing an attention mechanism, the focus is on abnormal objects in the scene and generative prediction is performed on severely occluded areas to output complete object 3D bounding boxes and uncertainty assessment.

[0011] As a preferred technical solution of this application, in step two, the grasping scene is first abstracted into a topological graph, and the spatial relationship between the object and the environment and obstacles is inferred using a graph neural network. Then, the long-term benefits and risks of different grasping strategies are evaluated through parallel Monte Carlo tree search. Finally, the grasping force control parameters are optimized online based on real-time tactile sensor data.

[0012] As a preferred technical solution of this application, in step three, a tightly coupled vision-to-inertial-to-laser SLAM technology is first used to construct a dynamic three-dimensional grid map containing semantic information. Then, a long short-term memory network is used to predict the future multi-step trajectory of dynamic obstacles. Finally, the temporal elastic band algorithm is combined to perform spatiotemporal joint optimization of the preliminary planned path to generate a safe, smooth and time-optimal grasping path.

[0013] In summary, this application includes at least the following beneficial technical effects of the AI-driven bomb disposal robot grasping method and system: This application, through the deep integration of multimodal perception, intelligent decision-making, and adaptive control, brings significant benefits. The system utilizes a Transformer-based visual model to achieve pixel-level accurate identification and 3D reconstruction even when explosives are severely obscured, providing high-precision geometric and physical information with uncertainty assessment for grasping. Based on this, the system can dynamically calculate the optimal grasping point and adaptive grasping force curve and generate a smooth, interference-resistant, collision-free trajectory through an improved planning algorithm. The digital twin module can complete tens of thousands of Monte Carlo simulations within seconds, quantifying risks in advance. Finally, the AI ​​decision-making is made transparent through an augmented reality interface, achieving efficient collaboration between human experience and machine intelligence, ensuring unprecedented safety, adaptability, and reliability in the grasping of high-risk explosives. Attached Figure Description

[0014] Figure 1 This is the overall system architecture diagram of this application; Figure 2 This is a flowchart of the bomb disposal robot grasping method of this application. Detailed Implementation

[0015] The following is in conjunction with the appendix Figure 1-2 This application will be described in further detail.

[0016] See Figure 1-2 The AI-driven bomb disposal robot grasping system includes: The intelligent recognition module is used to analyze the environmental data collected by the robot in real time using a large AI model. The intelligent recognition module is used to identify the shape, material and location features of explosives. The intelligent recognition module includes a visual analysis unit, a semantic segmentation unit and a 3D reconstruction unit. The visual analysis unit is used to extract multi-scale features from the video stream captured by the camera. The semantic segmentation unit uses a pre-trained model based on the Transformer architecture to accurately locate the outline of the explosive at the pixel level and predictively complete the occluded parts. The 3D reconstruction unit is used to generate a high-precision solid geometric model and centroid distribution of the target object based on point cloud data. In this application, the visual analysis unit adopts an architecture called "multi-scale feature pyramid network" to process the camera video stream in parallel and extract the local texture details and global shape context of the explosive. The semantic segmentation unit uses a Transformer-based pre-trained model. This model, through a self-attention mechanism, can understand the global dependencies of all pixels in the image. Thus, even if the explosive is occluded by more than 50%, it can accurately predict the reasonable contour of the occluded part based on the structural features of the visible part, achieving pixel-level segmentation accuracy. The 3D reconstruction unit further tightly couples and registers the 2D segmentation results with the LiDAR point cloud. Through iterative nearest-point algorithm and surface reconstruction algorithm, a high-precision 3D voxel model with uncertainty assessment is generated. This model can not only display the geometry, but also estimate the centroid position of the object through the point cloud density distribution, providing crucial input for grasping stability calculation.

[0017] The adaptive grasping strategy module dynamically generates grasping path, grasping angle, and grasping force parameters based on the recognition results. The module includes a grasping point optimization unit, a force control parameter generation unit, and an obstacle avoidance planning unit. The grasping point optimization unit calculates the optimal set of grasping contact points and evaluates grasping stability metrics based on the explosive's geometric characteristics, estimated center of mass, and physical stability. The force control parameter generation unit dynamically generates an adaptive grasping force curve by combining material recognition results and historical torque sensor data, and through a combination of table lookup and online learning. The obstacle avoidance planning unit generates a collision-free, smooth grasping trajectory based on a real-time updated voxel map and employs a fusion planning method combining an improved A* algorithm and a fast random tree algorithm. The adaptive grasping strategy module communicates with the intelligent recognition module. In this application, after receiving the 3D model output by the recognition module, the gripping point optimization unit discretizes it into thousands of potential contact point candidates. Then, a physics-based simulator is used to quickly calculate the force closure exponent and disturbance resistance capability for each gripping point combination, ultimately outputting a Pareto optimal solution set to achieve a balance between stability and obstacle avoidance capability. The force control parameter generation unit maintains a database of gripping force-slip relationship as prior knowledge, containing common materials and their surface properties. During the gripping process, a high-resolution torque sensor at the fingertip monitors minute slip signals in real time and uses an online recursive least squares algorithm to dynamically adjust the gripping force curve, enabling the system to adapt to changes in unknown materials or surface conditions and achieve adaptive gripping. The obstacle avoidance planning unit improves the traditional RRT* algorithm by combining it with model predictive control. While expanding the random tree, it continuously optimizes the trajectory in real time, ensuring that the generated path is not only collision-free but also smooth, energy-optimal, and predictively avoids suddenly entering dynamic obstacles.

[0018] The real-time environment perception module is used to monitor the interaction between the robot hand and the target object and environmental changes in real time through multi-sensor fusion technology. The real-time environment perception module includes a multi-sensor fusion unit, a dynamic obstacle tracking unit, and a contact state monitoring unit. The multi-sensor fusion unit uses a Kalman filter and deep learning fusion algorithm to simultaneously process data from the vision, tactile, acoustic, and inertial measurement units and construct a unified environmental state estimate. The dynamic obstacle tracking unit detects and predicts the trajectory of moving obstacles and calculates the collision risk probability based on target detection and optical flow. The contact state monitoring unit uses a high-precision torque sensor array to provide real-time feedback on torque changes during the grasping process. The real-time environment perception module is communicatively connected to the adaptive grasping strategy module. In this application, the multi-sensor fusion unit designs a hierarchical fusion strategy: the bottom layer uses Kalman filtering to preprocess data from homogeneous sensors, the middle layer uses a lightweight convolutional neural network to extract and correlate features from heterogeneous data such as vision, touch, and acoustics, and the top layer uses a gated recurrent unit network to perform temporal fusion of all features to output a unified and reliable environmental state estimate. The dynamic obstacle tracking unit combines advanced target detection models such as YOLOv7 with dense optical flow, which not only outlines moving obstacles in real time, but also predicts their trajectory in the next few seconds using a Kalman filter and calculates the collision risk level based on the trajectory intersection probability. The contact state monitoring unit uses a flexible torque sensor array deployed on the robot's fingertips and palm to collect six-dimensional force / torque data, which can sense micron-level slippage or rolling during the grasping process in real time and provide feedback for force control closed loop.

[0019] The grasping control module is used to control the robot hand to perform grasping actions based on the parameters output by the adaptive grasping strategy module and the feedback signals from the real-time environment perception module. The grasping control module is communicatively connected to the real-time environment perception module. The grasping control module receives the planned trajectory and force curve from the adaptive strategy module as feedforward instructions, and simultaneously receives tactile and visual feedback from the real-time sensing module as closed-loop correction signals. The core controller of the grasping control module adopts an impedance control strategy, but unlike fixed-parameter impedance control, its impedance parameters can be dynamically adjusted according to the grasping stage and the perceived environmental stiffness. For example, low stiffness is used when approaching the object to ensure safety, stiffness is increased at the moment of contact to suppress vibration, and damping is adjusted according to the material after stable grasping to prevent slippage. This effectively copes with the uncertainties and disturbances in the grasping process, ensuring that the grasping of fragile and easily deformable explosives is both stable and does not cause secondary damage.

[0020] The simulation verification and human-machine collaboration module is used to perform digital twin simulation verification of the strategy before physical grasping and realize human-machine collaborative decision-making through an augmented reality interface. The simulation verification and human-machine collaboration module includes a digital twin unit and an augmented reality interaction unit. The digital twin unit constructs a virtual scene synchronized with the real environment in real time based on the physics engine and performs Monte Carlo simulations of millions of times on the grasping strategy to be executed, and evaluates the grasping success rate and risk in a probabilistic form. The augmented reality interaction unit is used to overlay and display the grasping points, planned paths and force control parameters recommended by the AI ​​system on the operator's control interface and highlight the risk areas in a visual way. The operator has the final decision-making power and can modify or confirm the AI ​​plan. The simulation verification and human-machine collaboration module is communicatively connected to the grasping control module. In this application, the digital twin unit first simulates the dynamic characteristics of the robot, explosives, and environment with high fidelity through a physics engine. The key lies in the introduction of parameter perturbation and sensor noise models. Subsequently, the system performs tens of thousands of parallel "ultra-real-time" Monte Carlo simulations in the twin scene to test the proposed grasping strategy. Each simulation randomly changes the parameters, and finally outputs quantitative indicators such as the probability of successful grasping, maximum stress distribution, and estimated slip risk in the form of statistical distribution. The augmented reality interaction unit then presents these results intuitively: the operator can see the AI-recommended grasping points highlighted on the 3D model of the explosives through AR glasses or a display screen. The planned path is superimposed on the real video screen as a colored virtual trajectory, and risk areas are marked with semi-transparent red warning covers. The operator can directly drag and fine-tune the grasping points or paths in the AR interface with gestures or controllers. The system will re-simulate and update the risk indicators in real time. This interaction mode organically combines the computing power of AI with human experience and judgment, ensuring the reliability of the final decision.

[0021] The AI-driven bomb disposal robot grasping method includes the following steps: Step 1: The multimodal sensor data at the bomb disposal site is fused and analyzed using an AI big model to identify the type, shape, material, location, and attitude of the explosives. In Step 1, a multi-task learning framework is first used to process feature extraction, instance segmentation, and anomaly detection tasks in parallel. Then, by introducing an attention mechanism, the focus is on abnormal objects in the scene and generative prediction is performed on severely occluded areas, outputting complete 3D bounding boxes of objects and uncertainty assessment. The system first runs a multi-task model based on the Swim Transformer architecture. This model shares common features extracted by the backbone network and completes three sub-tasks in parallel: feature extraction, instance segmentation, and anomaly detection. Subsequently, the system introduces a cross-attention mechanism module, which calculates the similarity between the features of each region in the scene and the key features in the prior knowledge base of "explosives". This allows the system to automatically focus on high-risk targets with abnormal shapes, special materials, or concealed locations. For severely occluded objects, the system enables a conditional variational autoencoder. This model, after training, can generate multiple reasonable 3D structural hypotheses of the occluded parts based on the features of the visible parts and output the uncertainty score of each hypothesis. Finally, a 3D bounding box with uncertainty boundaries is formed, providing rich contextual information rather than a single deterministic result for subsequent strategy formulation.

[0022] Step two: Based on the identified multidimensional features and real-time environmental parameters, an adaptive grasping strategy containing grasping points, paths, angles, and force curves is generated through a policy network driven by a reinforcement learning algorithm. In step two, the grasping scene is first abstracted into a topological graph, and the spatial relationship between the object and the environment and obstacles is inferred using a graph neural network. Then, the long-term benefits and risks of different grasping strategies are evaluated through parallel Monte Carlo tree search. Finally, the grasping force control parameters are optimized online based on real-time tactile sensor data. In this application, the system first abstracts the robot, the target explosive, surrounding obstacles, and their spatial relationships into a topological graph, where nodes represent entities and edges represent spatial relationships. Then, a graph neural network is used to perform message passing and node feature updates on the entire graph, thereby inferring complex multi-body interaction relationships. Based on the embedding vectors output by the GNN, the policy network generates a set of preliminary grasping policy candidates. Subsequently, the system initiates parallelized Monte Carlo tree search to simulate the long-term consequences of executing these policies in parallel in a digital twin environment, quickly assessing their long-term benefits and risks. Finally, before actual grasping, the system preloads historical data from the tactile sensors and fine-tunes the initial parameters of the grasping force curve through an online optimizer to ensure that it can contact the object with near-optimal force from the beginning.

[0023] Step 3: Construct a dynamic voxel map using multi-sensor fusion data, track obstacles in real time, and use a model predictive control algorithm to optimize the grasping path planning. In step 3, firstly, a tightly coupled vision-inertial-laser SLAM technology is used to construct a dynamic 3D grid map containing semantic information. Then, a long short-term memory network is used to predict the future multi-step trajectory of dynamic obstacles. Finally, the temporal elastic band algorithm is combined to perform spatiotemporal joint optimization of the preliminary planned path to generate a safe, smooth, and time-optimal grasping path. In this application, the system first runs a tightly coupled vision-inertial-laser SLAM algorithm. This algorithm constructs a dynamic voxel map with semantic labels in real time by jointly optimizing visual feature points, IMU pre-integration constraints, and laser point cloud matching, and can accurately estimate the robot's pose. For dynamic obstacles in the map, the system uses a long short-term memory network to predict the set of possible trajectories for the next few steps based on their historical motion sequences. Finally, path planning is not completed in one step, but adopts a model predictive control framework combined with a time elastic band algorithm for rolling optimization. The TEB algorithm not only optimizes the spatial shape of the path, but also optimizes the time sequence of the robot's movement along the path, thereby generating a smooth and time-optimal grasping path that maintains a safe distance from the trajectory of dynamic obstacles in all four dimensions of time and space.

[0024] Step four: Through a force feedback closed loop based on impedance control, the posture and force of the gripping mechanism are adjusted in real time to ensure the smoothness, stability and safety of the gripping action. In this application, during the initial grasping stage, the controller primarily uses position control to drive the robot hand to precisely locate the target grasping point along the planned trajectory. When the fingertip sensor detects contact with the object, the controller immediately switches to a force-priority impedance control mode. The target impedance in the impedance model is adaptively set according to the identified object material. For example, when grasping fragile paper explosives, high damping and low stiffness are set to make the robot hand exhibit "soft" characteristics and avoid impact. When grasping smooth metal shells, the stiffness is appropriately increased to ensure the stability of the grasp. Throughout the process, a high-frequency six-dimensional torque sensor provides real-time feedback on the contact force / torque of the fingertip, which is compared with the desired force curve. The torque output of each joint is dynamically adjusted by the PID controller, forming an adaptive force feedback closed loop that can effectively suppress disturbances and absorb uncertainties, ensuring smooth, stable, and safe grasping.

[0025] Step 5: Before physical execution, the capture strategy is verified by ultra-real-time simulation in the digital twin system, and the optimal strategy is submitted to the operator for confirmation and execution through the AR interface. In this application, before physical execution, the entire grasping strategy is imported into a high-fidelity digital twin system. This system, based on a high-performance physics engine, simulates various uncertainties in the real world by injecting a large number of random perturbations. Subsequently, the system uses parallel computing technology to complete tens of thousands of repeated executions of the strategy within seconds. Each simulation is conducted under slightly different initial conditions. After the simulation, the system statistically analyzes the number of successful grasps, the maximum stress value, the probability of slippage, etc., and finally evaluates the reliability of the strategy in the form of probability distribution. These results, along with the AI's recommended scheme, are intuitively superimposed on the operator's real field of vision through an augmented reality interactive unit: the grasping point is displayed with a green halo, the path is a blue virtual trajectory, and high-risk areas are highlighted with semi-transparent red. The operator can view, zoom, and even directly "drag" the virtual grasping point in the AR interface for fine-tuning using gestures. The system will re-simulate and update the risk indicators in real time. Finally, the operator makes a decision to confirm or modify based on full knowledge, realizing the seamless integration and collaboration of artificial intelligence's super computing power and human expert experience judgment.

[0026] The above are all preferred embodiments of this application, and are not intended to limit the scope of protection of this application. Therefore, all equivalent changes made in accordance with the structure, shape and principle of this application should be covered within the scope of protection of this application.

Claims

1. An AI-driven bomb disposal robot grasping system, characterized by: include: The intelligent recognition module is used to analyze the environmental data collected by the robot in real time through an AI model, and is used to identify the shape, material and location characteristics of explosives. An adaptive grasping strategy module is used to dynamically generate grasping path, grasping angle and grasping force parameters based on the recognition results. The adaptive grasping strategy module is communicatively connected to the intelligent recognition module. A real-time environment perception module is used to monitor the interaction between the robot hand and the target object and environmental changes in real time through multi-sensor fusion technology. The real-time environment perception module is communicatively connected to the adaptive grasping strategy module. The grasping control module is used to control the robot hand to perform grasping actions based on the parameters output by the adaptive grasping strategy module and the feedback signal from the real-time environment perception module. The grasping control module is communicatively connected to the real-time environment perception module. The simulation verification and human-machine collaboration module is used to perform digital twin simulation verification of the strategy before physical grasping and realize human-machine collaborative decision-making through augmented reality interface. The simulation verification and human-machine collaboration module is communicatively connected to the grasping control module.

2. The AI-driven large-scale model-driven bomb disposal robot grasping system according to claim 1, characterized in that: The intelligent recognition module includes a visual analysis unit, a semantic segmentation unit, and a 3D reconstruction unit. The visual analysis unit is used to extract multi-scale features from the video stream captured by the camera. The semantic segmentation unit uses a pre-trained model based on the Transformer architecture to accurately locate the outline of the explosive at the pixel level and predictively complete the occluded parts. The 3D reconstruction unit is used to generate a high-precision solid geometric model and centroid distribution of the target object based on point cloud data.

3. The AI-driven large-scale model-driven bomb disposal robot grasping system according to claim 1, characterized in that: The adaptive grasping strategy module includes a grasping point optimization unit, a force control parameter generation unit, and an obstacle avoidance planning unit. The grasping point optimization unit is used to calculate the optimal set of grasping contact points and evaluate the grasping stability metric based on the geometric characteristics, estimated centroid, and physical stability of the explosive. The force control parameter generation unit is used to dynamically generate an adaptive grasping force curve by combining material identification results and historical data from torque sensors and by combining table lookup with online learning. The obstacle avoidance planning unit generates a collision-free smooth grasping trajectory based on a real-time updated voxel map and by using a fusion planning of an improved A* algorithm and a fast random tree algorithm.

4. The AI-driven large-scale model-driven bomb disposal robot grasping system according to claim 1, characterized in that: The real-time environment perception module includes a multi-sensor fusion unit, a dynamic obstacle tracking unit, and a contact state monitoring unit. The multi-sensor fusion unit uses a Kalman filter and deep learning fusion algorithm to simultaneously process data from the vision, tactile, acoustic, and inertial measurement units and construct a unified environmental state estimate. The dynamic obstacle tracking unit detects and predicts the trajectory of moving obstacles and calculates the probability of collision risk based on target detection and optical flow. The contact state monitoring unit uses a high-precision torque sensor array to provide real-time feedback on torque changes during the grasping process.

5. The AI-driven large-scale model-driven bomb disposal robot grasping system according to claim 1, characterized in that: The simulation verification and human-machine collaboration module includes a digital twin unit and an augmented reality interaction unit. The digital twin unit constructs a virtual scene synchronized with the real environment in real time based on a physics engine and performs Monte Carlo simulations of the grasping strategy to be executed millions of times to evaluate the grasping success rate and risk in a probabilistic manner. The augmented reality interaction unit is used to overlay and display the grasping points, planned paths and force control parameters recommended by the AI ​​system on the operator's control interface and highlight the risk areas in a visual manner. The operator has the final decision-making power and can modify or confirm the AI ​​plan.

6. An AI-driven large-scale model-based explosive ordnance disposal robot grasping method, as described in any one of claims 1-5, characterized in that: Includes the following steps: Step 1: Use AI big data models to fuse and analyze multimodal sensor data at the bomb disposal site to identify multidimensional features such as the type, shape, material, location, and attitude of explosives. Step 2: Based on the identified multidimensional features and real-time environmental parameters, an adaptive grasping strategy containing grasping points, paths, angles, and force curves is generated through a policy network driven by a reinforcement learning algorithm. Step 3: Construct a dynamic voxel map using multi-sensor fusion data, track obstacles in real time, and use model predictive control algorithms to optimize the grabbing path planning. Step four: Through a force feedback closed loop based on impedance control, the posture and force of the gripping mechanism are adjusted in real time to ensure the smoothness, stability and safety of the gripping action. Step 5: Before physical execution, the capture strategy is verified by ultra-real-time simulation in the digital twin system, and the optimal strategy is submitted to the operator for confirmation and execution through the AR interface.

7. The AI-driven bomb disposal robot grasping method according to claim 6, characterized in that: In step one, a multi-task learning framework is first used to process feature extraction, instance segmentation and anomaly detection tasks in parallel. Then, by introducing an attention mechanism, the focus is on abnormal objects in the scene and generative prediction is performed on severely occluded areas, outputting complete 3D bounding boxes of objects and uncertainty assessment.

8. The AI-driven bomb disposal robot grasping method according to claim 6, characterized in that: In step two, the grasping scene is first abstracted into a topological graph. Graph neural networks are used to infer the spatial relationships between objects and the environment and obstacles. Parallel Monte Carlo tree search is then used to evaluate the long-term benefits and risks of different grasping strategies. Finally, based on real-time tactile sensor data, the grasping force control parameters are optimized online.

9. The AI-driven bomb disposal robot grasping method according to claim 6, characterized in that: In step three, a tightly coupled vision-to-inertial-to-laser SLAM technology is first used to construct a dynamic 3D grid map containing semantic information. Then, a long short-term memory network is used to predict the future multi-step trajectory of dynamic obstacles. Finally, the temporal elastic band algorithm is combined to perform spatiotemporal joint optimization of the preliminary planned path to generate a safe, smooth and time-optimal grasping path.

Images