Battery assembly robot grabbing method based on deep learning

Abstract

The application belongs to the technical field of intelligent manufacturing and industrial robots, and particularly relates to a battery assembling robot grabbing method based on deep learning, which comprises the following steps: constructing a physical support atlas based on three-dimensional point cloud data, calculating metastable support indexes of each battery to evaluate collapse risks; combining robot kinematics and battery physical properties to calculate dynamic shearing safety indexes to prevent grabbed objects from slipping in high-speed motion; evaluating global exposure gains by simulating grabbing actions, optimizing grabbing sequences, and avoiding causing subsequent grabbing difficulties; and through a multi-target optimization strategy, the optimal grabbing target that can maximize the convenience of subsequent operations is selected under the premise of ensuring physical stability and dynamic safety. The application improves the safety, efficiency and intelligent level of the battery grabbing process.

Description

Technical Field

[0001] This invention relates to the field of intelligent manufacturing and industrial robot technology. More specifically, this invention relates to a deep learning-based gripping method for battery assembly robots. Background Technology

[0002] In the battery module production process for industries such as new energy vehicles, industrial robots need to accurately and stably grasp and assemble disordered stacked square batteries from deep frames. The level of automation and operational stability of this step directly affects the production efficiency and product quality of the entire production line.

[0003] However, in complex scenarios with tightly stacked deep frames, there are intricate physical contacts and force transmission relationships between batteries. Traditional robotic grasping methods, such as strategies that select the highest point based solely on visual height information, are often unsuitable. This is because the visually highest point may not be the physically most suitable object to grasp. If the robot mistakenly grasps a critical battery that serves as a support for the lower layer, it can cause the upper batteries to lose balance and collapse or slip, potentially damaging the battery electrodes and even causing equipment jamming, severely impacting production.

[0004] To address these issues, existing technologies have begun to employ grasping and planning methods based on 3D vision and physical models. These methods typically construct a support relationship map between objects by analyzing point cloud data and then assess the impact of removing any object on overall stability. For example, the local pressure is assessed by calculating the ratio of the total mass of the upper objects supported by each object to the contact area, thereby determining the fragility of the support relationship. However, in battery stacking scenarios, these methods have significant drawbacks: their stability assessment models are often overly simplified, failing to fully consider complex physical factors such as static friction between batteries and the potential energy risk brought about by stacking height. This results in inaccurate predictions of collapse risk, making it difficult to meet the high reliability requirements of industrial environments. Summary of the Invention

[0005] To address the technical problem of poor robot grasping performance during battery assembly, this invention provides a deep learning-based battery assembly robot grasping method, comprising:

[0006] Acquire 3D point cloud data of the stacked scene and identify individual batteries; construct a physical support map containing each battery; based on the physical support map, obtain the metastable support index of each battery. The metastable support index is positively correlated with the total mass of the upper-layer batteries supported by the battery and the relative layer height of the upper-layer batteries relative to the battery, and negatively correlated with the contact support area of ​​the upper-layer batteries on the battery; obtain the dynamic shear safety index of each battery, and screen candidate batteries with a metastable support index less than a set threshold and a dynamic shear safety index greater than a set safety margin; the dynamic shear safety index is the ratio of the contact support area between the suction cup and the corresponding battery. The difference between the maximum static friction force between the pool surfaces and the tangential shear force generated by the robot's motion; the maximum static friction force is positively correlated with the theoretical suction force and negatively correlated with the peeling resultant force along the normal direction of the suction cup; the peeling resultant force includes the normal acceleration component and the gravity component along the normal direction; the candidate batteries are virtually grasped, a virtual residual scene is generated, and the global exposure gain is calculated; the global exposure gain is positively correlated with the increased visible area of ​​adjacent batteries after grasping and negatively correlated with the indentation volume generated in the virtual residual scene; among the candidate batteries, the battery with the largest global exposure gain is selected as the optimal grasping target, and the robot is driven to perform the grasping.

[0007] This invention optimizes the grasping sequence by comprehensively constructing a physical support map to assess static collapse risk, calculating a dynamic shear safety index to assess slippage risk during high-speed movement, and simulating grasping to assess global exposure gain, thus establishing a multi-objective optimization strategy. This invention not only avoids the stacking collapse problem caused by neglecting physical support relationships in traditional methods, but also reduces the risk of battery slippage during high-speed robot movement and avoids subsequent grasping difficulties caused by improper current grasping. Therefore, while ensuring physical stability and dynamic safety, it improves the overall safety, efficiency, and intelligence level of the battery grasping process.

[0008] Preferably, the construction of the physical support map includes:

[0009] For any two batteries, perform geometric analysis. If the horizontal projection area of ​​the upper battery on the lower battery is greater than the set projection overlap threshold, and the minimum vertical distance between them is less than the set vertical contact tolerance, then establish a support edge between them from the lower to the upper layer.

[0010] This invention constructs a physical support map by setting two explicit geometric thresholds: the proportion of horizontal projected area and the minimum vertical distance. This enables more accurate and reliable identification of the actual physical support relationship between batteries, effectively eliminating interference from visually adjacent batteries that have no actual load-bearing relationship. This provides a more accurate basis for subsequent stability analysis and improves the reliability of metastable support index calculation.

[0011] Preferably, the metastable support index is also positively correlated with the static friction coefficient corrected by the tilt angle of the corresponding battery; the tilt angle is the angle between the normal vector of the upper surface of the battery and the vertical direction.

[0012] Preferably, the metastable support index further includes a risk penalty term that is exponentially related to the relative hierarchical height. ,in Indicates the floor height penalty coefficient; Indicates the first The upper support battery is relative to the first The relative hierarchical height of each battery.

[0013] Preferably, the dynamic shear safety index satisfies the expression:

[0014] ;

[0015] In the formula, Indicates the first The dynamic shear safety index of each battery; This represents the coefficient of static friction between the battery surface and the suction cup; Indicates theoretical suction force; Indicates the first The mass of each battery; Represents gravitational acceleration; Indicates the first The angle between the normal to the battery surface and the direction of gravity; Indicates normal acceleration; This indicates tangential acceleration.

[0016] This invention measures the competitive relationship between the maximum static friction force provided by the suction cup and the tangential shear force generated by the robot's motion. It can predict the risk of battery slippage under specific acceleration in real time, thereby incorporating dynamic stability into the grasping decision and avoiding the problem of the grasped object slipping or being thrown off due to high-speed motion.

[0017] Preferably, the method further includes performing a deceleration optimization cycle when the dynamic shear safety index of any candidate battery is not greater than the set safety margin: reducing the normal acceleration and tangential acceleration by a set step size, and recalculating the dynamic shear safety index until it is greater than the set safety margin or the acceleration is lower than the set minimum efficiency threshold.

[0018] Preferably, the global exposure gain satisfies the expression:

[0019] ;

[0020] In the formula, Indicates the capture of the first Global exposure gain after one battery; Indicates the first The set of neighboring cells of each cell; Indicates the removal of the first After the battery, its neighbor The newly added surface area that can be covered by vertical projection; Indicates the indentation penalty coefficient; Indicates the first Each battery corresponds to a projection area of ​​the virtual remaining scene; Indicates the first Virtual remaining scene depth map corresponding to each battery; Indicates the first Smooth depth map corresponding to each battery; This represents the newly added concave volume in the virtual remaining scene.

[0021] This invention achieves an assessment of the long-term impact of grasping actions by balancing the increased visible area after grasping with the volume of the resulting depression. It can guide the robot to select a grasping target that is not only optimal at present but also creates more favorable conditions for subsequent operations, avoiding the creation of deep pits that are difficult to reach and improving the smoothness and efficiency of the overall grasping task.

[0022] Preferably, the calculation of the depression volume includes:

[0023] A morphological closing operation is performed on the depth map of the virtual remaining scene to obtain a smoothed depth map; the depth difference between the smoothed depth map and the depth map of the virtual remaining scene in the corresponding battery projection area is integrated to obtain the concave volume.

[0024] This invention utilizes image processing technology to statistically analyze the volume of dead zones such as deep pits or narrow slits caused by the grasping process, which are detrimental to subsequent operations. This provides a reliable basis for calculating the global exposure gain, making the optimization of the grasping sequence more accurate and effective.

[0025] Preferably, the morphological closing operation is performed using a structuring element, the size of which is set to 1.2 times the diameter of the robot suction device.

[0026] Preferably, the acquisition of the optimal capture target further includes:

[0027] When using a multi-head suction device, search for battery combinations that satisfy the suction device's geometric constraints among the candidate batteries; calculate the total global exposure gain of each battery combination; and select the battery combination with the largest total gain as the optimal gripping target.

[0028] The beneficial effects of this invention are as follows: by constructing a physical support map and calculating the metastable support index, the static collapse risk of stacked objects can be predicted and avoided; the calculation of the dynamic shear safety index avoids the dynamic safety problem of the grasped object slipping during the robot's high-speed movement; by evaluating the global exposure gain, the grasping sequence can be pre-optimized to avoid falling into subsequent operational dilemmas; by comprehensively considering physical stability, dynamic safety and long-term operational convenience, the intelligence and reliability of robot grasping are improved. Attached Figure Description

[0029] Figure 1 This is a flowchart illustrating a battery assembly robot grasping method based on deep learning in this invention.

[0030] Figure 2 This is a schematic diagram illustrating the physical support structure.

[0031] Figure 3 It is a schematic diagram of a multi-dimensional feature space view for implementing the final crawling strategy. Detailed Implementation

[0032] This invention discloses a deep learning-based grasping method for a battery assembly robot, referring to... Figure 1 This includes steps S1-S4:

[0033] S1: Construct a physical support map based on 3D point cloud data and obtain the metastable support index of each battery.

[0034] It should be noted that in the deep-frame stacking scenario of battery module production lines, batteries, taking square aluminum-cased cells as an example, are typically stacked in a disordered yet compact manner. Due to the physical contact and force transmission between the batteries, a strategy based solely on the highest point based on visual height often fails. If a battery acting as a bottom support is grabbed, it can cause the upper batteries to lose balance and collapse or slip, resulting in electrode damage or jamming. Therefore, it is necessary to look beyond the visual appearance to uncover the physical support relationships between the batteries, analyze the load level of each battery in the stacked structure, and thus identify the key support nodes that can cause changes in other batteries. Based on this, this invention constructs a physical support map and calculates a metastable support index.

[0035] Specifically, a physical support map is constructed based on 3D point cloud data, and the metastable support index of each battery is obtained, including:

[0036] A 3D point cloud of the battery stacking scene is acquired, and several independent batteries are identified using a pre-trained instance segmentation network. For example, the instance segmentation network is Mask R-CNN, which is existing technology and will not be described in detail here.

[0037] It should be noted that visual overlap does not imply physical support; for a lack of support, both horizontal projection overlap and vertical close contact must be met simultaneously.

[0038] For any two batteries, a spatial geometric analysis is performed. If the horizontal projected area of ​​the upper battery on the lower battery is greater than the projection overlap threshold, and the minimum vertical distance between them is less than the vertical contact tolerance, then a directed support edge is established between them, pointing from the lower battery to the upper battery, thus constructing a physical support map for the entire field. For example, the projection overlap threshold is 0.3, and the vertical contact tolerance is... This means that at least 30% of the bottom area of ​​the upper battery layer is projected onto the lower battery layer, and the vertical gap between the upper and lower batteries must be less than [a certain value]. Only then can it be considered a valid contact.

[0039] It's important to note that to accurately calculate collapse risk, a comprehensive evaluation index is needed, incorporating load mass, contact area, and layer height. In physical stacking, the bottom layer of batteries often bears the weight of all the batteries above; the smaller the contact area, the greater the local pressure, and the less stable the support; furthermore, the higher the layer a battery is in, the greater its potential energy, and the stronger the impact damage caused by instability. Therefore, the support index should be directly proportional to the load mass and inversely proportional to the contact area. Simultaneously, considering that the collapse of higher-level batteries can trigger a chain reaction, the contribution of layer height to risk should increase exponentially.

[0040] During the offline calibration phase of the system, the static friction coefficient between batteries was obtained through friction experiments on battery casings of different materials; historical stacking data was collected to statistically analyze the relationship between stacking height and collapse probability, and regression fitting was performed to derive the layer height penalty coefficient, denoted as . For example, when the battery surface is rough or the stacking is relatively stable The value is 0.5 when the battery surface is smooth or the stacking is loose. The value is set to 1.0 to increase the penalty for higher-level loads. It should be noted that the regression fitting method, such as logistic regression, is existing technology and will not be elaborated upon here.

[0041] Let a small area constant be defined as . Exemplary .

[0042] For any battery, obtain the set of upper-level supporting batteries, which is the set of all upper-level batteries directly or indirectly supported by the battery in the physical support map. Use a plane fitting algorithm to fit the upper surface plane of the battery and obtain the normal vector of this plane. Calculate the angle between the normal vector and the vertical axis of the world coordinate system, denoted as the tilt angle of the battery. For any upper-level supporting battery, based on the physical support map, use a graph traversal algorithm, setting the battery as the root node, and search upwards in the directed graph. The shortest path length from the upper-level supporting battery to the root node is denoted as the relative level height of the upper-level supporting battery relative to the battery. The plane fitting algorithm is the RANSAC random sampling consensus algorithm, and the graph traversal algorithm is a breadth-first search algorithm; both are existing technologies and will not be elaborated upon here.

[0043] The metastable support exponent of any battery satisfies the following expression:

[0044] ;

[0045] In the formula, Indicates the first Metastability support index of individual batteries; Indicates the first The upper support battery assembly of each battery; Indicates the first The first battery The mass of the upper support battery; Indicates the static friction coefficient between batteries; Indicates the first The first battery The tilt angle of the upper supporting battery; , Indicates the k-th battery, the... The first battery A set of data points supporting the upper-level battery; It is the intersection symbol; Indicates the first The battery and the first The contact support area of ​​the upper supporting battery; Represents a small area constant; Represents the natural exponential function; Indicates the floor height penalty coefficient; Indicates the first The upper support battery is relative to the first The relative hierarchy height of each battery; Represents the normalization function; It represents the acceleration due to gravity.

[0046] In the formula, This represents the local pressure load of the i-th upper supporting battery on the k-th battery. The larger this value is, the greater the pressure per unit area and the more fragile the support relationship. A correction for slippage risk due to tilting is introduced; the larger the tilt angle, the larger the correction factor; exponential term This represents the potential energy risk amplification factor, indicating that the risk increases exponentially with the increase of relative floor height. By combining mechanical load and energy potential, the systemic collapse risk that may be triggered by removing the k-th battery is demonstrated.

[0047] It should be noted that, as Figure 2 This is a schematic diagram of the physical support structure. The background is a raw image of the battery stack. Each circular node in the diagram represents an identified independent battery cell. The blue arrows connecting the nodes represent the inferred physical support relationship. The arrows point from the bottom supporting battery to the supported upper battery. The color of the nodes corresponds to the metastable support index. The redder the color, the heavier the load on the upper layer and the higher the layer, and the greater the risk of collapse if the battery is removed. The lighter the color, the lighter the load and the more stable the structure.

[0048] Thus, the metastable support index of each battery was obtained.

[0049] S2: Based on the robot's motion parameters and the battery's physical properties, obtain the dynamic shear safety index of each battery and obtain a set of candidate batteries that meet the dynamic stability requirements.

[0050] It should be noted that the battery surface is usually covered with an insulating film or residual oil film, making it relatively smooth. Furthermore, industrial robots typically operate at high acceleration to meet cycle time requirements. Under these conditions, vacuum suction alone is often insufficient to counteract the inertial forces generated by the robot's acceleration and deceleration, leading to lateral shearing slippage or even detachment of the battery from the suction cup. Therefore, a complete dynamic model incorporating friction, inertial force, gravity, and vacuum suction is needed to anticipate slippage risks during the motion planning stage and ensure dynamic stability during the grasping process. Based on this, this invention constructs a dynamic shearing safety criterion to obtain a set of candidate batteries that meet the requirements of dynamic stability.

[0051] Specifically, based on robot motion parameters and battery physical properties, the dynamic shear safety index of each battery is obtained, and a candidate battery set that satisfies dynamic stability is obtained, including:

[0052] Obtain the robot's preset motion trajectory parameters, including normal acceleration and tangential acceleration.

[0053] All batteries with metastable support indices below a preset safety threshold are used as the initial screening set. It should be noted that an index below the preset safety threshold indicates that the corresponding battery is not a critical support node; for example, the safety threshold is 0.8.

[0054] It should be noted that, in order to construct a reliable criterion, it is necessary to compare the anti-slip capability with the slip-inducing external force. During the grasping process, the battery is subjected to the suction force of the suction cup, as well as the inertial force from gravity and the robot's motion. Slippage occurs when the tangential component of the net external force exceeds the maximum static friction. The anti-slip capability is determined by the maximum static friction, which is equal to the surface friction coefficient multiplied by the actual normal force, and is also the suction force minus the normal inertial force. The slip-inducing external force is determined by the vector sum of gravity and the tangential inertial force. Only when the former is greater than the latter is the grasping safe. Therefore, this invention obtains a dynamic shear safety index based on the comparative relationship between the anti-slip capability and the slip-inducing external force.

[0055] During the offline calibration phase of the system, the static friction coefficient between the insulating film on the battery surface and the rubber of the suction cup was determined by dragging experiment; the rated negative pressure value of the vacuum generator, the effective contact area of ​​the suction cup, and the system leakage efficiency factor were obtained.

[0056] The theoretical suction force is the product of the rated negative pressure of the vacuum generator, the effective contact area of ​​the suction cup, and the system leakage efficiency factor.

[0057] The dynamic shear safety index of any battery satisfies the following expression:

[0058] ;

[0059] In the formula, Indicates the first The dynamic shear safety index of each battery; This represents the coefficient of static friction between the battery surface and the suction cup; Indicates theoretical suction force; Indicates the first The mass of each battery; Represents gravitational acceleration; Indicates the first The angle between the normal to the battery surface and the direction of gravity; Indicates normal acceleration; This indicates tangential acceleration.

[0060] In the formula, This represents the normal peeling acceleration, indicating the combined effect of the gravitational component and the inertial force, which attempts to pull the battery away from the suction cup. This reflects the tendency to detach due to the gravitational component and normal inertial force, which weakens the effective positive pressure of the suction cup on the battery, thereby reducing the upper limit of friction. The effective positive pressure is the remaining force that the suction cup actually applies to the battery surface after overcoming the peeling tendency; This indicates the maximum available static friction force that the suction cup interface can provide under the current high acceleration conditions. This represents the actual tangential shear resultant force acting on the contact surface, including the tangential inertial force and the tangential component of gravity. This criterion intuitively reflects the competing relationship of forces; only when the available frictional force can completely cover the shear resultant force with a margin can the shear resultant force be effectively controlled. In this case, crawling is dynamically secure.

[0061] Set a safety redundancy threshold, denoted as For example, under normal operating conditions, Under conditions with high safety requirements, such as human-machine collaboration... .

[0062] Batteries with a dynamic shear safety index greater than the safety redundancy threshold are marked as candidate battery sets.

[0063] It should be noted that if the dynamic shear safety index is less than or equal to the safety redundancy threshold, the system enters an iterative deceleration optimization loop: with a preset step size, such as 10% of the current acceleration, the normal acceleration and tangential acceleration in the motion plan are reduced, and the dynamic shear safety index is recalculated until the criterion is met or the acceleration is lower than the minimum efficiency threshold set by the system, at which point it is determined that it cannot be grabbed.

[0064] At this point, the candidate battery set for each battery has been obtained.

[0065] S3: Construct a virtual remaining scene based on simulated grabbing actions of candidate battery sets, and obtain the global exposure gain of each grabbing action.

[0066] It should be noted that grasping is a continuous sequential decision-making process. The current grasping action directly determines the stacking shape of the remaining batteries. If a greedy strategy is adopted, focusing only on immediate convenience, it is easy to create deep pits, narrow gaps, or islands in the stack. As a result, although subsequent batteries are visually visible, they cannot be grasped because the grippers or suction device cannot reach them, leading to a degradation in accessibility. Therefore, this invention introduces a multi-step look-ahead mechanism, prioritizing actions that make the remaining stacked surface more flat and open, increase convexity, and reduce concave dead corners.

[0067] Specifically, a virtual remaining scene is constructed based on simulated grabbing actions of candidate battery sets, and the global exposure gain of each grabbing action is obtained, including:

[0068] For any battery in the candidate battery set, query the physical support map to obtain all spatially adjacent batteries of the battery; perform virtual removal in memory, update the pixel values ​​of the corresponding battery region in the original scene depth map to the depth values ​​of the lower background, and generate a virtual remaining scene depth map.

[0069] It should be noted that, in order to analyze the merits of stacking configurations, this invention constructs an evaluation function that rewards a smooth surface and penalizes deep pits and depressions. After a good grasping action, more surrounding batteries should be exposed, increasing their area, and there should be no deep pits that are difficult to penetrate in the scene, reducing depressions. Therefore, the global gain is divided into a positive gain term and a negative penalty term. The positive gain is determined by the increased visible area of ​​adjacent batteries, while the negative penalty is determined by the non-convexity of the scene.

[0070] Define a structuring element for morphological closing operations, the size of which is set according to the physical size of the suction cup. For example, the structuring unit is 1.2 times the diameter of the suction cup, thereby simulating the physical size constraints of the suction cup.

[0071] A morphological closing operation is performed on the virtual remaining scene depth map to generate a smooth depth map. It should be noted that this operation fills in all pits and slits smaller than the structuring element.

[0072] A depression penalty coefficient is set to balance area gains and terrain risks, denoted as . For example, when a task is in the initial cleanup phase and the target needs to be exposed quickly, Take the smaller value, such as 10 When the task is in the deep-frame cleanup phase, and blind spots need to be avoided, Take the larger value, such as 50 .

[0073] The global exposure gain of any battery satisfies the expression:

[0074] ;

[0075] In the formula, Indicates the capture of the first Global exposure gain after one battery; Indicates the first The set of neighboring cells of each cell; Indicates the removal of the first After the battery, its neighbor The newly added surface area that can be covered by vertical projection; Indicates the indentation penalty coefficient; Indicates the first Each battery corresponds to a projection area of ​​the virtual remaining scene; Indicates the first Virtual remaining scene depth map corresponding to each battery; Indicates the first Smooth depth map corresponding to each battery; This represents the newly added concave volume in the virtual remaining scene.

[0076] In the formula, This indicates the field of view brought about by the grabbing action. The larger the value, the more batteries can be exposed after the battery is grabbed, so that those obstructions that are pressing on other batteries can be grabbed earlier. The property of filling deep pits using morphological closing operations This reflects the volume of blind spots that no suction device can reach. The larger this value is, the more rugged the remaining scene is and the more difficult it is to perform subsequent operations. By subtracting this integral term, the algorithm can effectively suppress actions that, although revealing the area, create terrain traps.

[0077] At this point, the global exposure gain for each grabbing action has been obtained.

[0078] S4: Perform multi-objective optimization based on metastable support index and global exposure gain, and obtain the optimal capture strategy.

[0079] It should be noted that the final grasping decision needs to strike a balance between physical safety and global accessibility. Pursuing only exposure gain may trigger dangerous batteries with high metastable support exponents, while excessive conservatism may lead to low grasping efficiency. Therefore, this invention constructs an evaluation system that integrates multiple objectives to maximize the operability of the remaining space while ensuring that collapse and slippage do not occur. Based on this, this invention performs multi-objective optimization and obtains the optimal grasping strategy.

[0080] Specifically, multi-objective optimization is performed based on the metastable support index and global exposure gain to obtain the optimal capture strategy, including:

[0081] The selection of the optimal crawling strategy satisfies the following logic:

[0082] Hard constraint filtering eliminates all batteries with metastable support indices greater than a safety threshold, ensuring that the dynamic shear safety index of the remaining batteries is greater than a safety redundancy threshold. Soft constraint sorting selects the battery with the largest global exposure gain from the remaining set after hard constraint filtering as the optimal grasping target. Multi-head collaboration, if using a multi-head suction device, searches the remaining set for battery combinations that satisfy the device's geometric constraints, calculates the sum of the global exposure gains of all batteries in the combination, and selects the combination with the largest sum as the optimal grasping target. For example, the suction device's geometric constraints include suction cup spacing constraints and height difference tolerance constraints. It should be noted that a dynamic shear safety index less than or equal to the safety redundancy threshold indicates a risk of slippage, and a metastable support index greater than the safety threshold indicates a risk of collapse; such batteries are strictly prohibited from being grasped.

[0083] It should be noted that, as Figure 3To provide a multi-dimensional feature space view for the final grasping strategy, the horizontal axis represents the metastable support index, and the vertical axis represents the dynamic shear safety index. The size and color intensity of the scatter points represent the global exposure gain. The red background area in the figure represents the collapse risk zone where the metastable support index exceeds the limit, and the orange background area represents the slippage risk zone where the dynamic shear safety index is insufficient. Batteries falling into these two areas are filtered out by hard constraints. In the remaining safe white area, the system selects the battery with the largest global exposure gain as the optimal target based on soft constraints, which is the area marked by the red bold circle, and drives the robot to perform the grasping.

[0084] At this point, the optimal grasping strategy was obtained, and the robot was driven to perform the grasping task.

[0085] While various embodiments of the invention have been shown and described in this specification, it will be apparent to those skilled in the art that such embodiments are provided by way of example only. Many modifications, alterations, and alternatives will occur to those skilled in the art without departing from the spirit and essence of the invention.

Claims

1. A battery assembly robot grasping method based on deep learning, characterized in that, include: Acquire 3D point cloud data of the stacked scene and identify individual batteries; Construct a physical support map containing each battery; Based on the physical support map, the metastable support index of each battery is obtained. The metastable support index is positively correlated with the total mass of the upper battery supported by the battery and the relative height of the upper battery to the battery, and negatively correlated with the contact support area of ​​the upper battery on the battery. The dynamic shear safety index of each battery is obtained, and candidate batteries with a metastable support index less than a set threshold and a dynamic shear safety index greater than a set safety margin are selected. The dynamic shearing safety index is the difference between the maximum static friction force between the suction cup and the corresponding battery surface and the tangential shearing force generated by the robot's movement; the maximum static friction force is positively correlated with the theoretical suction force and negatively correlated with the peeling resultant force along the normal direction of the suction cup; the peeling resultant force includes the normal acceleration component and the gravity component along the normal direction. The candidate batteries are virtually grasped to generate a virtual remaining scene and calculate the global exposure gain. The global exposure gain is positively correlated with the increased visible area of ​​neighboring batteries after grasping and negatively correlated with the indentation volume generated in the virtual remaining scene. Among the candidate batteries, the battery with the largest global exposure gain is selected as the optimal grasping target, and the robot is driven to perform the grasping.

2. The battery assembly robot grasping method based on deep learning according to claim 1, characterized in that, The construction of the physical support map includes: For any two batteries, perform geometric analysis. If the horizontal projection area of ​​the upper battery on the lower battery is greater than the set projection overlap threshold, and the minimum vertical distance between them is less than the set vertical contact tolerance, then establish a support edge between them from the lower to the upper layer.

3. The battery assembly robot grasping method based on deep learning according to claim 1, characterized in that, The metastable support index is also positively correlated with the static friction coefficient corrected by the tilt angle of the corresponding battery; the tilt angle is the angle between the normal vector of the upper surface of the battery and the vertical direction.

4. The battery assembly robot grasping method based on deep learning according to claim 1, characterized in that, The metastable support index also includes a risk penalty term that is exponentially related to the relative hierarchical height. ,in Indicates the floor height penalty coefficient; Indicates the first The upper support battery is relative to the first The relative hierarchical height of each battery.

5. The battery assembly robot grasping method based on deep learning according to claim 1, characterized in that, The dynamic shear safety index satisfies the expression: ； In the formula, Indicates the first The dynamic shear safety index of each battery; This represents the coefficient of static friction between the battery surface and the suction cup; Indicates theoretical suction force; Indicates the first The mass of each battery; Represents gravitational acceleration; Indicates the first The angle between the normal to the battery surface and the direction of gravity; Indicates normal acceleration; This indicates tangential acceleration.

6. The battery assembly robot grasping method based on deep learning according to claim 1, characterized in that, It also includes performing a deceleration optimization cycle when the dynamic shear safety index of any candidate battery is not greater than the set safety margin: reducing the normal acceleration and tangential acceleration by a set step size, and recalculating the dynamic shear safety index until it is greater than the set safety margin or the acceleration is lower than the set minimum efficiency threshold.

7. The battery assembly robot grasping method based on deep learning according to claim 1, characterized in that, The global exposure gain satisfies the expression: ； In the formula, Indicates the capture of the first Global exposure gain after one battery; Indicates the first The set of neighboring cells of each cell; Indicates the removal of the first After the battery, its neighbor The newly added surface area that can be covered by vertical projection; Indicates the indentation penalty coefficient; Indicates the first Each battery corresponds to a projection area of ​​the virtual remaining scene; Indicates the first Virtual remaining scene depth map corresponding to each battery; Indicates the first Smooth depth map corresponding to each battery; This represents the newly added concave volume in the virtual remaining scene.

8. The battery assembly robot grasping method based on deep learning according to claim 1, characterized in that, The calculation of the depression volume includes: A morphological closing operation is performed on the depth map of the virtual remaining scene to obtain a smoothed depth map; the depth difference between the smoothed depth map and the depth map of the virtual remaining scene in the corresponding battery projection area is integrated to obtain the concave volume.

9. A battery assembly robot grasping method based on deep learning according to claim 8, characterized in that, The morphological closing operation is performed using a structuring element whose size is set to 1.2 times the diameter of the robot suction device.

10. The battery assembly robot grasping method based on deep learning according to claim 1, characterized in that, The acquisition of the optimal crawling target also includes: When using a multi-head suction device, search for battery combinations that satisfy the suction device's geometric constraints among the candidate batteries; calculate the total global exposure gain of each battery combination; and select the battery combination with the largest total gain as the optimal gripping target.

Images