Metal object grasping posture perception and planning method for redundant robot

Abstract

The application discloses a metal object grasping posture perception and planning method for a redundant robot. In the perception layer, a physical heuristic coloring model 3D Gaussian sputtering technology is used for three-dimensional reconstruction, and a rasterization method is used to render and generate a virtual depth map to compensate for perception defects caused by extreme lighting and metal characteristics. In the decision-making layer, a scoring function is designed to fuse force closure, normal consistency, geometric center, collision avoidance and task direction preference, and the optimal grasping posture is selected through comprehensive evaluation. In the motion planning layer, a quadratic programming control framework based on mixed force / speed operability is constructed for the redundant robot, which optimizes the motion and force exertion ability of the robot in the carrying task under the premise of triple constraints of joint position, speed and acceleration, and realizes balanced distribution of joint torque. The application is suitable for high-precision and high-reliability grasping and carrying of high-reflectivity and heavy-load metal components in extreme environments such as the moon and space stations.

Description

Technical Field

[0001] This invention belongs to the field of robot vision perception and intelligent control, specifically relating to a method for perceiving and planning the posture of grasping metal objects for redundant robots. It is particularly suitable for high-precision and high-stability grasping and handling of highly reflective metal components in extreme lighting environments such as lunar base construction. Background Technology

[0002] With the development of deep space exploration and extraterrestrial base construction, automated operations on the lunar surface by redundant lunar robots have become crucial. In scenarios such as lunar base construction, a large number of metal components (such as aluminum alloy and titanium alloy structural parts) need to be handled and assembled. However, the extreme lighting conditions on the lunar surface, coupled with the strong specular reflection of metal surfaces, result in numerous noise, holes, and even large areas of missing data in depth images acquired by depth cameras, posing a significant challenge to the robot's visual perception system. At the grasping planning level, during the handling of heavy metal components, there is a contradiction between the system's motion capability and force application capability. Mechanically, one or more joints of the robotic arm often approach their torque output limit, leading to the failure of the grasping and handling task. Therefore, there is an urgent need for a perception planning grasping method that can collaboratively solve the perception distortion of metal objects under extreme lighting conditions and balance motion and force application capabilities when handling heavy metal objects, in order to improve the reliability, safety, and efficiency of robot operations in complex environments such as the lunar surface. Summary of the Invention

[0003] This invention aims to overcome the shortcomings of existing technologies, such as inaccurate perception of metal objects under complex lighting conditions, difficulty in generating grasping postures, and neglect of the dynamic constraints of the actuator in heavy-duty grasping planning, making it difficult to achieve a balance between motion and force. This invention provides a full-link grasping method from perception to planning to achieve safe, stable, and efficient grasping of highly reflective metal objects.

[0004] Technical solution:

[0005] A method for posture perception and planning in grasping metal objects for redundant robots includes the following steps:

[0006] Step 1: Use a camera to capture multi-view RGB images around the metal object, and use a shading function that integrates diffuse, specular, and residual terms to perform 3D Gaussian sputtering reconstruction of the scene; based on the reconstructed scene, generate a virtual depth map through rasterization rendering.

[0007] Step 2: Input the RGB image and virtual depth map into the grasping pose generation network to obtain an initial grasping pose candidate set; design a hybrid physical metric scoring function. The function includes at least the force closure metric S. d Geometric center measurement S gNormal consistency measure S f Collision avoidance metric S c and task orientation preference metric S t Each candidate pose is evaluated based on five weighted components, and the pose with the highest score is selected as the optimal grasping pose.

[0008] Step 3: Construct a quadratic programming control framework based on hybrid force / velocity maneuverability. The objective function of this framework includes an end-effector trajectory tracking term, a joint motion regularization term, and a hybrid maneuverability optimization term. The framework is solved under multiple constraints of joint position, velocity, and acceleration to generate control commands that drive the redundant robot to move to the optimal grasping posture in an optimized configuration and perform grasping.

[0009] Furthermore, in step 1, the coloring function is expressed as:

[0010] ,

[0011] Where γ is the gamma tone mapping function, c d represents the diffuse color of the Gaussian sphere, and s is the specular gloss level defined for the sphere. This indicates that the sphere is in the direction Direct specular illumination on the surface, n is the normal to the Gaussian sphere, ρ represents the roughness of the sphere, c r The residual color is represented by a third-order spherical harmonic function, and ⊙ represents element-wise multiplication.

[0012] Furthermore, in step 1, the 3D Gaussian sputtering reconstruction involves: combining the parameters of the Gaussian primitives optimized by differentiable rendering, and a Gaussian adaptive density control method to reconstruct a Gaussian 3D scene. The parameters of the Gaussian primitives optimized by differentiable rendering include the center position x. c Covariance Σ, Opacity α, Coloring coefficient c.

[0013] Furthermore, in step 1, generating a virtual depth map through rasterization rendering specifically includes: mapping the three-dimensional ray-Gaussian intersection problem to affine space through local projection transformation for calculation, for pixels on the image plane. Its final depth value Take the median of the depth values ​​contributed by all Gaussian volumes covering the pixel; the final depth value is expressed as:

[0014] ,

[0015] In the formula, Indicates the covered pixels The set of all three-dimensional Gaussian solids, for the k-th Gaussian solid, Let its center be the depth in the camera coordinate system. Let its center be the projection coordinates on the image plane. It is due to its three-dimensional covariance The constant vector derived from the camera extrinsic parameters through local affine projection. This represents the median operator.

[0016] Furthermore, in step 2, the geometric center measurement S g Based on the weighted center point coordinates P of the Gaussian elements that make up the grasped object gc Calculations are performed: ,in, It is the three-dimensional center point vector of the i-th Gaussian element, and N is the total number of Gaussian elements. ;

[0017] , where P cl P cr These are the left contact point and the right contact point, respectively.

[0018] Furthermore, the normal consistency score S f Represented as: ,

[0019] Where, n i For two contact points P cl P cr The normal vector n′ is represented in ray space as: ,in J is a vector determined by the Gaussian parameters and camera extrinsic parameters. a The local affine transformation matrix transforms the normal vector in ray space to the camera coordinate system, v. a Let be the antipodal direction vector of the gripper. This indicates the inner product operation.

[0020] Furthermore, the task orientation preference metric S t for: ,

[0021] Among them, v b It is the direction in which the gripper approaches the object, v e It represents the expected direction of the force applied in the task, and β is the sharpening index, used to enhance the discriminative power of the direction selection. When β>1, the function is more lenient towards small deviations close to the task direction, while penalizing large deviations more severely.

[0022] Furthermore, the collision deflection metric is: ,

[0023] in, , Indicates the ideal grab distance. Parameters for controlling the acceptable range.

[0024] Furthermore, in step 3, the hybrid force / velocity operability gradient function is defined as: , and These are adjustable state-related weighting coefficients used to balance the emphasis on velocity and force performance; and These represent speed / force operability, respectively.

[0025] Furthermore, in step 3, the objective function of the quadratic programming control framework is:

[0026] ,

[0027] Among them, the system's generalized joint variables Generalized joint velocity: w t w v w m These are the weighting coefficients for trajectory tracking, joint velocity regularization, and force operability, respectively. For the rotation Jacobian matrix, ω max and ω min The maximum and minimum permissible angular velocities.

[0028] Furthermore, in step 3, in the quadratic programming solution, the constraint q of each joint j j,min q j,max The limits are determined by the joint position limits, velocity limits, and acceleration limits, specifically:

[0029] ,

[0030] In the formula, j=1, …, 10, t is the control period, and q j,min q j,max For the joint position limit, v j,min v j,max The maximum speed limit of the joint is... Maximum acceleration limit of joint, q j This indicates the current joint position.

[0031] Compared with the prior art, the present invention has the following advantages:

[0032] 1. Reconstruction is performed using 3D Gaussian sputtering based on a shading enhancement model. High-quality virtual depth maps and surface normal maps are generated through rasterized rendering, fundamentally overcoming the problem of missing depth information caused by metal reflection under extreme lighting conditions.

[0033] 2. A hybrid physical metric is proposed, which not only considers the traditional principle of force closure, but also integrates the metrics of contact normal consistency, geometric center, contact collision avoidance, and task preference. This allows for the selection of metal object grasping postures that better meet the actual physical interaction and task requirements, significantly improving the grasping success rate and stability.

[0034] 3. At the motion planning layer, the coordinated optimization of trajectory tracking and force / velocity operability is realized. While ensuring the accuracy of end-effector trajectory tracking, the motion and force application capabilities of the robotic arm to perform specific tasks are improved, ensuring that the robotic arm is in a good kinematic and mechanical state during grasping and subsequent handling, and improving the operational robustness of the system under heavy load conditions.

[0035] 4. This invention discloses a robot perception and planning system for metal objects under complex lighting conditions, realizing a complete technology chain from anti-interference perception to intelligent grasping and planning, which greatly reduces the need for manual intervention and is suitable for complex scenarios with extremely high requirements for autonomy and reliability, such as lunar base construction. Attached Figure Description

[0036] Figure 1 This is a schematic diagram of the technical route of one embodiment of the present invention;

[0037] Figure 2 This is a schematic diagram of the intersection point of a ray and a Gaussian body according to an embodiment of the present invention;

[0038] Figure 3 This is a demonstration diagram of the object grasping posture according to an embodiment of the present invention;

[0039] Figure 4 This is a schematic diagram of the definition of a redundant robot coordinate system according to an embodiment of the present invention. Detailed Implementation

[0040] To make the objectives, technical solutions, and advantages of this invention clearer, the following detailed description, in conjunction with the accompanying drawings and embodiments, further illustrates the invention. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of the invention.

[0041] To address the severe lack of depth perception data due to specular reflection from highly reflective objects such as metal components in complex lighting environments like lunar base construction, and the control failures that arise from neglecting joint torque constraints and motion stability in existing grasping planning methods during heavy object handling, this embodiment discloses a posture perception and planning method for grasping metal objects in redundant robots. Figure 1As shown, firstly, at the perception level, based on multi-view RGB images, high-fidelity 3D reconstruction is performed using a physically inspired shading model and 3D Gaussian sputtering technique. A high-quality virtual depth map is generated using rasterization, fundamentally compensating for perception defects caused by extreme lighting and the characteristics of metallic objects. Secondly, at the decision-making level, a hybrid physical metric scoring function is designed, integrating force closure, normal consistency, geometric center, collision avoidance, and task direction preference, to comprehensively evaluate candidate grasping postures and select the optimal grasping posture that meets the task requirements. Finally, at the motion planning level, for redundant robots, a quadratic programming control framework based on hybrid force / velocity maneuverability is constructed. Under the premise of satisfying the triple constraints of joint position, velocity, and acceleration, the motion and force application capabilities of the redundant robot when performing handling tasks are optimized, achieving balanced distribution of joint torques to avoid exceeding limits and ensuring safe, stable, and efficient grasping execution.

[0042] Specifically, it includes the following steps:

[0043] Step 1: Gaussian sputtering 3D reconstruction based on shading enhancement representation.

[0044] An RGB image dataset was generated by taking pictures of a metallic object using a D415C camera carried by a redundant robot. The dataset was then processed using a structure-of-motion (SOR) method from the input image set {I}. t Generate a sparse point cloud P for a scene in the range |t=1, 2, ..., T}, and assign a value to each image I. t Estimating the initial camera pose ξ t The spherical harmonic color model used in standard 3DGS is abandoned. Based on this, a 3D Gaussian sputtering technique with shading enhancement representation is used for 3D scene reconstruction. A shading function is employed that decomposes color into three components: diffuse, specular, and residual.

[0045]

[0046] Where γ is the gamma tone mapping function, c d represents the diffuse color of the Gaussian sphere, and s is the specular gloss level defined for the sphere. This indicates that the sphere is in the direction Direct specular illumination on the surface, n is the normal to the Gaussian sphere, ρ represents the roughness of the sphere, c r The residual color is represented by a third-order spherical harmonic function, and ⊙ represents element-wise multiplication.

[0047] Combined with differentiable rendering to optimize the parameters of the Gaussian meta-element, including geometric properties: center position x c The high-fidelity Gaussian 3D scene is reconstructed using covariance Σ, opacity α, appearance attributes: shading coefficient c, and a Gaussian adaptive density control method.

[0048] Step 2: After performing 3D reconstruction using Step 1, based on the geometric principle of the intersection of light rays and Gaussian volumes, high-quality virtual depth maps and surface normal maps are generated through rasterization rendering, providing high-precision image input for subsequent capture generation.

[0049] In the camera coordinate system, consider a ray of light originating from the camera's central viewpoint o, with a direction of r, such as... Figure 2 As shown, its parameterized representation is:

[0050]

[0051] Where s is the distance from the point along the ray to the center of the camera. The ray passes through a 3D Gaussian volume distribution. When , its energy distribution can be modeled as a one-dimensional Gaussian function with respect to s:

[0052]

[0053] Where, x c Centered on Gauss, Let be the covariance matrix. The effective intersection point of a ray and a Gaussian body is defined as... The parameter s* corresponding to the point that takes the maximum value can be obtained by solving the following optimization problem in closed form:

[0054]

[0055] A local projection transformation is introduced to achieve efficient rasterization computation, mapping the 3D ray-Gaussian intersection problem to an affine space. In this ray space, the ray direction is normalized as follows: Gaussian center x c Mapped to The covariance matrix is ​​transformed accordingly as follows: Intersection parameter s ∗ It can be represented as:

[0056]

[0057] in Indicates the starting point of the projection corresponding to the current pixel. The projection coefficient vector associated with the Gaussian parameters:

[0058]

[0059] The depth value of this intersection point in the original camera coordinate system can be approximated as:

[0060]

[0061] in, It is a vector corresponding to a fixed Gaussian spot.

[0062] Under the assumption of local affine projection, the intersection points of all rays intersecting the same Gaussian volume form a plane. For pixels covered by the projected Gaussian distribution... The precise depth of each pixel is efficiently calculated using rasterization. Considering that multiple Gaussian bodies may overlap on the same ray, the median depth of each pixel is taken as the final depth value.

[0063]

[0064] in, Indicates the covered pixels The set of all three-dimensional Gaussian solids. For the k-th Gaussian solid, Let its center be the depth in the camera coordinate system. The projection coordinates of its center onto the image plane. It is due to its three-dimensional covariance This is a constant vector derived from the camera extrinsic parameters through local affine projection. This vector encodes the local geometric tilt information of the Gaussian volume within the pixel plane. This represents the median operator.

[0065] Step 3: Optimal grasping posture selection based on hybrid physical metrics.

[0066] The RGB image of the metal object captured by the camera carried by the redundant robot's end effector, the mask image, and the virtual depth map generated in the previous steps are input into the grasping pose generation network (GraspNet) to obtain an initial 6-DOF grasping pose candidate set. Each detected grasping posture includes rotation R j ∈SO(3) and grasping center t j ∈R 3 To select the optimal pose best suited for the actual task from the candidate set, a hybrid physics metric scoring function is designed for comprehensive evaluation. A demonstration of the grasping pose is shown below. Figure 3 As shown, the scoring function S evaluates each candidate pose based on five weighted components:

[0067]

[0068] Among them, S t S f S g S c S d These are based on force closure, geometric center, normal consistency, collision perturbation, and direction preference measures for task guidance.

[0069] Geometric center measurement: In practical grasping tasks, due to the difficulty in obtaining the center of gravity, this invention proposes a standard for geometric center point measurement, such as... Figure 3 As shown, the geometric center is determined by the center of the Gaussian elements that make up the grasping object. To more accurately reflect the contribution of each Gaussian element to the overall shape, a weighted average can be used, with the weights based on the opacity α of the elements. i and its covariance matrix Σ i The determined volume influence range, the coordinates P of the geometric center point gc It can be represented as:

[0070]

[0071] in, It is the three-dimensional center point vector of the i-th Gaussian element, and N is the total number of Gaussian elements. .

[0072] The closer the force direction between candidate points is to the geometric center of the object, the more stable the grip. Combining the geometric characteristics of the gripping posture, the two contact points are connected by the antipodal line. Finally, the Euclidean distance from the Gaussian geometric center point to the antipodal line is used as the geometric center score S. g :

[0073]

[0074] Where P cl P cr These are the left contact point and the right contact point, respectively.

[0075] Normal consistency metric: For the grasping task of a two-finger gripper, the consistency between the contact surface normal and the gripper normal directly affects the stability of the grasp. Normal consistency score S f :

[0076]

[0077] Where, n i For two contact points P cl P cr The normal vector n′ is represented in ray space as: ,in It is a vector determined by the Gaussian parameters and camera extrinsic parameters. J a This is a local affine transformation matrix that transforms the normal vectors in ray space to the camera coordinate system. a Let be the antipodal direction vector of the gripper. This indicates the inner product operation.

[0078] Collision perturbation metric: To prevent objects from colliding with the grippers due to excessive close contact, a collision perturbation score S is defined. c :

[0079]

[0080] in: , Indicates the ideal grab distance. Parameters for controlling the acceptable range.

[0081] Task orientation preference metric: Considering the direction of force application during actual task execution, task orientation preference is incorporated into the grasping evaluation.

[0082]

[0083] Among them, v b It is the direction in which the gripper approaches the object, v e It represents the expected direction of the force applied in the task, and β is the sharpening index (default value is 2), which is used to enhance the discriminative power of the direction selection. When β>1, the function is more lenient towards small deviations close to the task direction, while penalizing large deviations more severely.

[0084] All candidate grasping poses are scored according to the scoring function, and the grasping pose g with the highest score is selected. f This serves as the final grasping posture for the redundant robot.

[0085] Step 4: Grasping motion planning and execution based on force / speed operability optimization.

[0086] Based on the scoring criteria selection using the hybrid physical metrics in step 3, we obtained the optimal grasping posture g for reflective metallic objects at the task level. f For a redundant robot, there are infinitely many joint configurations that can make its end effector achieve the same target pose g. f Different configurations for grasping tasks correspond to completely different mechanical states of the robotic arm, which may lead to the robotic arm being in a strange configuration, joint over-limit, or low maneuverability, thus affecting the stability and success rate of grasping execution. This invention addresses the need for redundant robots to balance motion flexibility and force application capability in complex tasks by proposing a quadratic programming control framework based on hybrid force / velocity maneuverability. This framework can control the redundant robot to move safely, smoothly, and efficiently to the optimal grasping pose selected in step 3, while ensuring good mechanical performance during grasping and subsequent handling.

[0087] The forward kinematics of the redundant robot are derived jointly from the kinematic models of both the mobile platform and the robotic arm. For example... Figure 4 The diagram illustrates the definition of redundant robot coordinate systems, which includes the world coordinate system Σw, the base coordinate system Σb, the robot arm coordinate system Σm, the end effector coordinate system Σee, the camera coordinate system Σc, and the target object coordinate system Σo.

[0088] The positive kinematics of the robotic arm relative to Σm is:

[0089]

[0090] in, It is the attitude of the end effector in Σm; It is the generalized coordinate system of the robotic arm; To establish the transformation matrix between the joints of a seven-DOF robotic arm using the improved DH parameter method; This indicates the transformation from the flange of the 7th joint of the robotic arm to the end effector, where l is the tool length.

[0091] The target pose is defined as follows:

[0092]

[0093] The camera is fixedly mounted at the end of the robotic arm. Determined by hand-eye signification (eye on the hand), It is the pose of the grasped object in the camera coordinate system, output by the grasping pose generation system.

[0094] Under the pure rolling assumption, the kinematic model of the moving base is as follows:

[0095]

[0096] This represents the speed of the differential wheels. This is the base constraint matrix.

[0097] The generalized velocity of a robotic arm can be expressed in terms of joint velocities. The speed of the end effector is actually the differential of equation x(q) with respect to time. Combining this with the above equation, we can see that:

[0098]

[0099] in and These are the Jacobian matrices for the base and the robotic arm, respectively. This is the complete Jacobian matrix of the system.

[0100] To improve the dynamic performance of redundant robots in collaborative handling tasks, their configuration is optimized to maximize their motion and force transmission capabilities in task-related directions. This invention designs a hybrid force / velocity operability gradient function, which is achieved by weighted fusion of velocity operability gradient and force operability gradient:

[0101]

[0102] in, and These are adjustable state-related weighting coefficients used to balance the emphasis on velocity and force performance; and These represent speed / force operability, both along the unit direction n. v and n f Defined as:

[0103]

[0104]

[0105] in, Let be the position Jacobian matrix. The larger the value, the more likely it is to be along n. v The greater the flexibility of directional movement, The larger the value, the more likely it is to be along n. f It has the best directional force application capability.

[0106] The robotic arm's end effector uses a quaternion-based control method. A standard quaternion is defined as... ,in For scalar branches, For vector branches, satisfying Current position of the end effector and posture For the grasping posture of a metal object, given the desired position and the desired pose matrix Rotation matrix and The corresponding quaternion forms are respectively represented as and The attitude error between the two Combined quaternion representation is ,in, Therefore, the attitude motion error of the robotic arm's end effector can be defined as follows: In the formula It is a skew-symmetric operator.

[0107] Based on combined position and attitude control, the desired end-effector velocity is:

[0108]

[0109] in, and These are the desired terminal velocity and angular velocity vectors, respectively. and These represent the position control and attitude error control gains, respectively. and These represent the position error and attitude error vectors, respectively. .

[0110] To address the motion control problem of redundant robots, a single-layer quadratic programming framework is constructed. This framework couples the trajectory tracking task and the force maneuverability optimization task into the same objective function, utilizing the robot's redundant degrees of freedom to optimize its force / velocity maneuverability, thus avoiding the complexity of traditional hierarchical priority strategies. The objective function is designed as follows:

[0111]

[0112] Among them, the system's generalized joint variables Generalized joint velocity: w t w v w m These are the weighting coefficients for trajectory tracking, joint velocity regularization, and force operability, respectively. For the rotation Jacobian matrix, ω max and ω min The maximum and minimum permissible angular velocities.

[0113] To achieve safe and smooth motion, the allowable velocity range of each joint needs to be calculated in real time based on triple constraints of joint position, velocity, and acceleration. This paper adopts a hierarchical constraint fusion method to ensure that the generated motion commands do not exceed the physical limits of the robot. For each joint j, its upper and lower velocity limits are determined by the intersection of the following three factors:

[0114]

[0115] In the formula, j=1, …, 10, t is the control period, and q j,min q j,max For the joint position limit, v j,min v j,max The maximum speed limit of the joint is... Maximum acceleration limit of joint, q j This indicates the current joint position.

[0116] The single-layer QP framework achieves coordinated optimization of trajectory tracking and force operability, which improves the robotic arm's ability to perform specific tasks while ensuring the accuracy of the end-effector trajectory.

[0117] This invention is applicable to the field of robot vision perception and control, and is especially suitable for high-precision and high-reliability grasping and handling of highly reflective and heavy-duty metal components in extreme lighting environments such as the moon and space stations. It is of great significance for promoting the automated construction of extraterrestrial bases.

[0118] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for posture perception and planning in grasping metal objects for redundant robots, characterized in that, Includes the following steps: Step 1: Use a camera to capture multi-view RGB images around the metal object, and use a shading function that integrates diffuse, specular, and residual terms to perform 3D Gaussian sputtering reconstruction of the scene; based on the reconstructed scene, generate a virtual depth map through rasterization rendering. Step 2: Input the RGB image and virtual depth map into the grasping pose generation network to obtain an initial grasping pose candidate set; Design a hybrid physical metric scoring function The function includes at least the force closure metric S. d Geometric center measurement S g Normal consistency measure S f Collision avoidance metric S c and task orientation preference metric S t Each candidate pose is evaluated based on five weighted components, and the pose with the highest score is selected as the optimal grasping pose. Step 3: Construct a quadratic programming control framework based on hybrid force / velocity maneuverability. The objective function of this framework includes an end-effector trajectory tracking term, a joint motion regularization term, and a hybrid maneuverability optimization term. The framework is solved under multiple constraints of joint position, velocity, and acceleration to generate control commands that drive the redundant robot to move to the optimal grasping posture in an optimized configuration and perform grasping.

2. The method for metal object grasping posture perception and planning according to claim 1, characterized in that, In step 1, the coloring function is expressed as: ， Where γ is the gamma tone mapping function, c d represents the diffuse color of the Gaussian sphere, and s is the specular gloss level defined for the sphere. This indicates that the sphere is in the direction Direct specular illumination on the surface, n is the normal to the Gaussian sphere, ρ represents the roughness of the sphere, c r The residual color is represented by a third-order spherical harmonic function, and ⊙ represents element-wise multiplication.

3. The method for metal object grasping posture perception and planning according to claim 1, characterized in that, In step 1, the 3D Gaussian sputtering reconstruction involves: combining the parameters of the Gaussian primitives optimized by differentiable rendering with a Gaussian adaptive density control method to reconstruct a Gaussian 3D scene. The parameters of the Gaussian primitives optimized by differentiable rendering include the center position x. c Covariance Σ, Opacity α, Coloring coefficient c.

4. The method for metal object grasping posture perception and planning according to claim 1, characterized in that, In step 1, generating a virtual depth map through rasterization rendering specifically includes: mapping the three-dimensional ray-Gaussian intersection problem to affine space through local projection transformation for calculation, for pixels on the image plane. Its final depth value Take the median of the depth values ​​contributed by all Gaussian volumes covering the pixel; the final depth value is expressed as: ， In the formula, Indicates the covered pixels The set of all three-dimensional Gaussian solids, for the k-th Gaussian solid, Let its center be the depth in the camera coordinate system. Let its center be the projection coordinates on the image plane. It is due to its three-dimensional covariance The constant vector derived from the camera extrinsic parameters through local affine projection. This represents the median operator.

5. The method for metal object grasping posture perception and planning according to any one of claims 1-4, characterized in that, In step 2, the geometric center measurement S g Based on the weighted center point coordinates P of the Gaussian elements that make up the grasped object gc Calculations are performed: ， in, It is the three-dimensional center point vector of the i-th Gaussian element, and N is the total number of Gaussian elements. ; ， Among them, P cl P cr These are the left contact point and the right contact point, respectively.

6. The method for metal object grasping posture perception and planning according to any one of claims 1-4, characterized in that, The normal consistency score S f Represented as: , Where, n i For two contact points P cl P cr The normal vector n′ is represented in ray space as: ,in J is a vector determined by the Gaussian parameters and camera extrinsic parameters. a The local affine transformation matrix transforms the normal vector in ray space to the camera coordinate system, v. a Let be the antipodal direction vector of the gripper. This indicates the inner product operation.

7. The method for metal object grasping posture perception and planning according to any one of claims 1-4, characterized in that, The task orientation preference metric S t for: , Among them, v b It is the direction in which the gripper approaches the object, v e It is the expected direction of the force applied in the task, and β is the sharpening index, which is used to enhance the discrimination of the direction selection. When β>1, the function is more lenient to small deviations close to the task direction, and more severe in its punishment of large deviations. The collision deflection metric is: , in, , Indicates the ideal grab distance. Parameters for controlling the acceptable range.

8. The method for metal object grasping posture perception and planning according to claim 1, characterized in that, In step 3, the hybrid force / velocity operability gradient function is defined as: , and These are adjustable state-related weighting coefficients used to balance the emphasis on velocity and force performance; and These represent speed / force operability, respectively.

9. The method for metal object grasping posture perception and planning according to claim 1, characterized in that, In step 3, the objective function of the quadratic programming control framework is: ， Among them, the system's generalized joint variables Generalized joint velocity: w t w v w m These are the weighting coefficients for trajectory tracking, joint velocity regularization, and force operability, respectively. For the rotation Jacobian matrix, ω max and ω min The maximum and minimum permissible angular velocities.

10. The method for metal object grasping posture perception and planning according to claim 8 or 9, characterized in that, In step 3, in the quadratic programming solution, the constraint q of each joint j j,min q j,max The limits are determined by the joint position limits, velocity limits, and acceleration limits, specifically: ， In the formula, j = 1, …, 10, To control the period, q j,min q j,max For the joint position limit, v j,min v j,max This is the maximum speed limit of the joint. To limit the maximum acceleration of the joint, q j This indicates the current joint position.

Images