A robot pushing and grabbing collaborative work method and device based on active interactive perception

Abstract

The application discloses a robot pushing and grabbing collaborative work method and device based on active interactive perception, and the method comprises the following steps: constructing a three-dimensional scene point cloud of a work space and reversely deducing the position and posture of a workpiece to be grabbed, adopting feature twin mapping to determine a theoretically grabbable part of the workpiece to be grabbed and performing double evaluation of the theoretically grabbable part by a visual filter and a physical filter; constructing a local causal topology graph, calculating a comprehensive performance cost ratio of pushing obstacles, and determining an optimal pushing path; pushing the obstacles along the optimal pushing path and performing physical deadlock detection during the pushing process; vertically lifting and resetting the robot at the end of the pushing path, again constructing a three-dimensional scene point cloud of the work space to review the pushing termination state, and performing grabbing when it is confirmed that the workpiece to be grabbed can be directly grabbed after the obstacles are pushed. The robot pushing and grabbing collaborative work method can eliminate occlusion by active pushing and pushing interaction in a scattered and stacked scene, and then realizes accurate grabbing of a target part.

Description

Technical Field

[0001] This invention relates to the field of robotics, and in particular to a method and apparatus for collaborative robot pushing and grasping operations based on active interactive perception. Background Technology

[0002] With the rapid development of industrial automation technology, industrial robots are playing an increasingly important role in fields such as automobile manufacturing, 3C electronics assembly, and logistics warehousing. Among these applications, picking up randomly stacked parts from a bin is a typical scenario for industrial robots. In this scenario, the robot system is usually equipped with high-precision 3D machine vision equipment (such as structured light cameras, laser contour scanners, or ToF cameras). By acquiring point cloud data or depth images of the workpieces in the bin, and combining this with a pre-made CAD model for feature matching and pose estimation, the robot plans the motion trajectory of the robotic arm to complete the task of identifying and picking up specific workpieces.

[0003] In real-world unstructured work environments, parts within a bin are often stacked in multiple disordered layers, frequently resulting in target parts being partially or completely obscured by other parts. For such complex scenarios, current mainstream technologies typically employ a "top-priority grasping" strategy, where the vision system prioritizes locking onto parts located on the top layer of the bin with fully exposed surface features. When the vision system cannot directly identify a graspable target due to occlusion, existing production lines often use physical aids to alter the distribution of parts. For example, vibrating the entire bin with a vibratory feeder, or controlling a robotic arm carrying an end effector to enter the bin and perform stirring or tumbling actions along a preset trajectory, attempting to expose the underlying parts through physical disturbance. Furthermore, some cutting-edge industrial applications and academic research are exploring interactive perception-based work modes, allowing robots to apply a certain amount of force to objects in the environment to change their position even with limited visual information. Current implementations mainly rely on simple geometric rules or random point selection strategies, such as identifying the highest point of the stack for pushing, or removing obstacles along a fixed direction, thereby re-triggering the visual recognition process by changing the object's positional distribution to find potential grasping opportunities.

[0004] The aforementioned existing technologies have the following drawbacks: (1) Lack of targeted "blind" interaction leads to low work efficiency. Existing physical disturbance methods (such as vibratory feeders or random stirring by robotic arms) are usually global or random operations. The system lacks semantic understanding of the occlusion relationship between parts in the scene and cannot determine which obstacle is blocking the target part. Therefore, the robot often performs a lot of ineffective pushing actions, and even pushes parts that are originally in a graspable state into dead corners, resulting in extremely long work cycles, which is difficult to meet the needs of efficient production lines.

[0005] (2) Focusing only on object displacement while ignoring the geometric requirements of the "grasping cavity". Most current pushing strategies only have the single goal of "removing obstructions", that is, pushing the obstacle away from the current position, without fully considering the physical space (i.e., the grasping cavity) required by the end effector of the robotic arm (such as a two-finger gripper) when performing the grasping action. This often leads to an awkward situation in actual operation: although the obstacle has been pushed away a certain distance, its position still occupies the motion envelope space required for the gripper to close, so the robotic arm still cannot lower the gripper and a second or even multiple pushes are required. Summary of the Invention

[0006] The purpose of this invention is to provide a robot push-grab collaborative operation method and device based on active interactive perception, which can eliminate the obstruction of the workpiece to be grasped by obstacles through active pushing and interaction in unstructured scattered stacking scenarios, thereby achieving accurate grasping of the workpiece to be grasped.

[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows: This invention provides a robot push-grab collaborative operation method based on active interactive perception, comprising the following steps: Environmental understanding and accessibility assessment: Construct a 3D scene point cloud within the workspace and match it with a pre-stored workpiece model. Inversely infer the position and orientation of the workpiece to be grasped in the workspace. Call a preset grasping fingerprint database to perform feature twin mapping to determine the theoretically graspable parts of the workpiece to be grasped. The theoretically graspable parts are then evaluated by both visual and physical filters to confirm whether the workpiece to be grasped can be directly grasped. When the workpiece to be grasped cannot be directly grasped, an active interaction strategy is executed. The active interaction strategy includes generating the optimal pushing path, obstacle pushing, and steady-state grasping. Generate the optimal migration path: Construct a local causal topology graph of the target and obstacles to determine the set of obstacles. For each obstacle in the set of obstacles, traverse its discrete migration direction in the plane and calculate the overall efficiency cost ratio (Cost) of migrating each obstacle along the discrete migration direction. Determine the optimal migration path based on the migration direction corresponding to the minimum overall efficiency cost ratio (Cost). Obstacle removal: Move obstacles along the optimal path and perform physical deadlock detection during the process. When a physical deadlock is detected, a high-frequency perturbation strategy is triggered. If the high-frequency perturbation strategy can unlock the obstacle, the removal continues. If the high-frequency perturbation strategy cannot unlock the obstacle, the corresponding obstacle is marked as a physical deadlock node in the local causal topology graph and a suboptimal removal path is replanned for removal. Steady-state grasping: After the robot is vertically lifted at the end of the pushing path, it resets and reconstructs the 3D scene point cloud in the workspace to confirm whether it can directly grasp the workpiece after pushing the obstacle. If the obstacle blocking the workpiece has been removed and both the visual filter and the physical filter have passed the evaluation, the workpiece is grasped directly. Otherwise, the local causal topology graph of the target and obstacle is reconstructed with the current scene as the initial state.

[0008] Furthermore, based on any one or a combination of the aforementioned technical solutions, the condition for being able to inversely deduce the position and orientation of the workpiece to be grasped in the workspace after matching the 3D scene point cloud in the workspace with the pre-stored workpiece model is that the vision system can capture at least one local feature of the workpiece to be grasped.

[0009] Furthermore, following any one or a combination of the aforementioned technical solutions, the preset grasping fingerprint database refers to the set of key geometric surfaces predefined in the CAD model of the workpiece that are adapted to the stable gripping of the robot's end effector.

[0010] Furthermore, following any one or a combination of the aforementioned technical solutions, the feature twin mapping method is as follows: based on the position and orientation of the workpiece to be grasped in the workspace inverse inference, the grasping fingerprint region of the workpiece to be grasped in the grasping fingerprint database is virtually reprojected onto the current three-dimensional scene point cloud to determine the spatial coordinates of the theoretically graspable part of the workpiece to be grasped in the current state.

[0011] Furthermore, following any one or a combination of the aforementioned technical solutions, the evaluation using the visual filter refers to calculating the exposure α of the fingerprint of the workpiece to be grasped. Specifically, the method involves using a depth comparison method based on camera viewpoint projection to discretize the fingerprint surface after feature twin mapping into a uniform set of sampling points. The effective proportion of the fingerprint region of the workpiece to be grasped after virtual reprojection, which is not obscured by environmental obstacles, is calculated. The formula for calculating the exposure α is α=(S total -S occluded ) / S total ×100%, where S total S represents the theoretical total number of sampling points for fingerprint capture in physical space via feature twin mapping. occluded This represents the number of sampling points in the area that are obscured by environmental obstacles. The evaluation using the physical filter refers to virtual physical collision envelope detection. The specific method is as follows: load the virtual collision envelope box of the robot's end effector along the grasping and probing path, and detect whether the end effector entity will interfere with the surrounding environmental obstacles during the probing process.

[0012] Furthermore, following any one or a combination of the aforementioned technical solutions, when the exposure α is greater than a preset threshold and the virtual physical collision envelope detection result is without interference, it is determined that the workpiece to be grabbed can be directly grabbed; otherwise, an active interaction strategy is executed.

[0013] Furthermore, following any one or a combination of the aforementioned technical solutions, in the step of environmental understanding and accessibility assessment, if the constructed 3D scene point cloud fails to match the effective target features of the workpiece to be grasped after matching it with the pre-stored workpiece model, then all the workpieces to be grasped are in a state of full occlusion. The system determines that the current scene cannot be grasped and directly triggers a global blind state perturbation strategy to make the robot flip the workpieces or vibrating boxes in the workspace until the vision system captures at least one local feature of the workpiece to be grasped.

[0014] Furthermore, following any one or a combination of the aforementioned technical solutions, the method for constructing a local causal topology graph of the target and obstacles is as follows: taking the occluded or spatially restricted workpiece to be grasped as the root node, and taking the envelope of the grasping fingerprint tool path that occupies the workpiece in space and / or the set of environmental point clouds that directly cover the model of the workpiece to be grasped as interference nodes, and establishing directed edges from the upper object to the lower object based on the physical support relationship, thereby obtaining a local causal topology graph of the target and obstacles; the set of obstacles includes all interference nodes in the local causal topology graph.

[0015] Furthermore, following any one or a combination of the aforementioned technical solutions, in the step of generating the optimal moving path, based on the result of traversing the discrete moving directions of each obstacle in the obstacle set along its plane, it is necessary to calculate the overall efficiency cost ratio (Cost) for moving each obstacle along each discrete moving direction. The overall efficiency cost ratio (Cost) is calculated using the following formula: Cost = (w1·V cloud +w2·D clutter ) / (G cavity ), where V cloud V is the resistance term. cloud V is the estimated physical resistance based on the bounding box volume or projected area of ​​the obstacle point cloud. It represents the estimated physical resistance that needs to be overcome when performing a pushing operation on a specific obstacle. cloud The value range of D is 0-1; clutter For the environmental item, D clutter It is a quantitative assessment of the spatial margin ahead based on 3D scene point clouds, representing the degree of environmental congestion in the direction of movement, D. clutter The value range of G is 0-1; cavity For the gain term, G cavityThis refers to the effective gain of the gripping cavity of the workpiece to be gripped, where the gripping cavity refers to the physical space required by the robot's end effector to perform the gripping action. cavity The value range of w1 is 0-1; w1 and w2 are both weighting coefficients, w1+w2=1 and the value range of w1 and w2 is 0-1. When the workpiece in the workspace is a heavy, large workpiece, w1>w2; when the workpiece in the workspace is a light, small workpiece, w1<w2.

[0016] Furthermore, following any one or a combination of the aforementioned technical solutions, the V cloud The V is obtained by calculating the geometric features of the obstacle point cloud on the pushing support surface to equivalently evaluate the frictional resistance. cloud V is calculated using the following formula: cloud =(k·S proj ) / F max_safe , of which S proj S is the projected area of ​​the smallest bounding box of the obstacle point cloud on the moving plane. proj It is positively correlated with frictional resistance; k is a preset resistance mapping coefficient, and the value of k ranges from 0.1 to 10; F max_safe This is the preset maximum safe thrust threshold that the robot is allowed to output in this workspace; when V cloud When the value is greater than 1, the system determines that the pushing operation carries the risk of overload shutdown or serious damage to the workpiece, and sets the overall efficiency cost ratio (Cost) corresponding to the pushing direction to infinity and discards it.

[0017] Furthermore, following any one or a combination of the aforementioned technical solutions, the D clutter The following method is used to obtain the virtual displacement envelope corridor constructed by the system along the candidate displacement direction, where the environmental congestion level D is... clutter Defined as the proportion of space within the shifting envelope corridor that is occupied by other environmental obstacles, i.e., D clutter =V obs / V corridor , where V corridor V represents the theoretical total volume of the shifting envelope corridor constructed by the system. obs V represents the total volume of other environmental obstacles located within the shifting envelope corridor, detected by visual point cloud; the shifting envelope corridor refers to the spatial volume swept across by the obstacle to be shifted plus a preset safety margin. obs This can be equivalent to the number of obstacle point clouds or voxel grids within the shifting envelope corridor; if the D clutter When the value is greater than 0.9, it indicates that there is a dead angle ahead in the direction of movement. The overall performance cost ratio (Cost) corresponding to the direction of movement is set to infinity and eliminated.

[0018] Furthermore, following any one or a combination of the aforementioned technical solutions, the G... cavity The gain term G is obtained by loading the virtual motion envelope box required for the robot's end effector to complete the closing action into the 3D scene point cloud using the following method. cavity G is defined as the proportion of the physical interference volume within the virtual motion envelope that can be eliminated by the pushing operation. cavity =(V t0 -V t1 ) / V t0 , where V t0 The initial interference volume is the total volume of the virtual motion envelope box that the current obstacle point cloud intrudes into and occupies before the push operation; V t1 To predict the residual interference volume, which is the volume of the obstacle point cloud that remains within the virtual motion envelope after the system virtually simulates the obstacle in a specific direction and distance.

[0019] Furthermore, following any one or a combination of the aforementioned technical solutions, the method for determining the optimal moving path based on the moving direction corresponding to the minimum overall efficiency cost ratio (Cost) is as follows: After calculating and comparing the overall efficiency cost ratio (Cost) of each obstacle moving along each discrete moving direction, the moving direction corresponding to the minimum overall efficiency cost ratio (Cost) is selected as the optimal moving direction. Then, a virtual simulation is performed on the obstacles along the selected optimal moving direction to obtain the moving distance along the optimal moving direction. During the virtual simulation, when V... t1 The minimum translation amount corresponding to reducing the obstacle to no longer obstruct the robot's expected grasping and feeding path is the distance pushed along the optimal pushing direction; Preset overall performance cost threshold (Cost) max If the minimum overall performance cost is greater than the cost max This directly triggers the global blind state perturbation strategy and forces the system state machine to jump and reset, returning to the environmental understanding and reachability assessment steps.

[0020] Furthermore, in accordance with any or a combination of the aforementioned technical solutions, in the obstacle pushing step, when the robot starts the pushing operation and the distance between the robot end and the target obstacle is 5-10mm, the robot switches to the variable impedance control mode to push the target obstacle.

[0021] Furthermore, following any one or a combination of the aforementioned technical solutions, the method for detecting physical deadlock is as follows: a preset pushing resistance threshold F is used. safe During the migration process, the migration resistance F is continuously monitored. extThe dynamic relationship between the actual moving speed v of the end effector and the monitored pushing resistance F ext The surge exceeded the drag resistance threshold F. safe If the actual moving speed v of the end effector suddenly drops to 0 and remains so for 0.1-0.5s, it is determined that mechanical coupling or static friction lock-up has occurred.

[0022] Furthermore, following any one or a combination of the aforementioned technical solutions, the high-frequency perturbation strategy involves pausing linear feed and superimposing high-frequency, low-amplitude sinusoidal vibrations in the tangential or normal directions of the contact surface; when the high-frequency perturbation strategy is adopted, the pushing resistance F ext Recover to less than the pushing resistance threshold F safe If the actual moving speed v of the end effector recovers, it is determined to be unlocked, and the robot resumes pushing and continues to complete the pushing.

[0023] Furthermore, following any or a combination of the aforementioned technical solutions, in the steady-state grasping step, the method for grasping the workpiece to be grasped is as follows: the robot moves from its reset starting point to directly above the workpiece to be grasped and vertically downwards to perform the grasping action. At the instant the robot's end effector contacts the surface of the workpiece to be grasped, the controller rapidly attenuates the normal contact force in the Z-axis direction to 0. During the downward movement and closing of the robot's end effector, the robot's robotic arm maintains low impedance characteristics on the horizontal plane. These low impedance characteristics are achieved by dynamically adjusting the stiffness parameters in the Cartesian space impedance controller, setting the translational stiffness of the robot's robotic arm in both the X and Y axes to 50-500 N / m. After the end effector confirms successful grasping closure, a stiffness locking mode is adopted to move the workpiece to be grasped out of the workspace, completing the push-grab collaborative operation of the workpiece to be grasped.

[0024] According to another aspect of the present invention, a robot push-grab collaborative operation device based on active interactive perception is provided, which uses the robot push-grab collaborative operation method based on active interactive perception as described above to complete the pushing of obstacles in the workspace and the grasping of workpieces to be grasped.

[0025] The beneficial effects of the technical solution provided by this invention are as follows: (1) By constructing a local causal topology graph of target-obstacle, this invention can accurately locate the key obstacles that truly hinder grasping. This interactive strategy ensures that the robot only pushes the necessary obstacles, effectively avoiding pushing the workpiece that could originally be grasped into a dead corner, thereby improving the working rhythm and overall efficiency of the production line and overcoming the drawbacks of random stirring or blind vibration in traditional technology. (2) Unlike the traditional strategy that only aims to remove obstacles, the present invention introduces the index of "grabbing cavity gain" in the calculation of the overall efficiency cost ratio. It specifically calculates whether the physical space required for the end effector to close can be freed up after the push. The system follows the principle of "stop when the space is sufficient" and dynamically calculates the minimum push distance, which not only ensures the physical accessibility of the subsequent gripping action, but also maintains the relative stability of the workpiece distribution in the workspace to the maximum extent. (3) The present invention introduces physical deadlock detection during the obstacle pushing process. By identifying mechanical deadlock caused by abnormal increase in resistance in real time, such as the common hooking and jamming of complex workpieces, once physical deadlock is detected, the system will no longer blindly push hard, but will trigger a high-frequency micro-perturbation strategy. By superimposing micro vibrations, the contact state will be separated from the static friction boundary, which can effectively protect the robot from overload while smoothly unlocking and getting out of trouble. Attached Figure Description

[0026] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0027] Figure 1 This is a flowchart illustrating a robot push-grab collaborative operation method based on active interactive perception, provided as an exemplary embodiment of the present invention. Detailed Implementation

[0028] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0029] One embodiment of the present invention provides a robot push-grab collaborative operation device based on active interactive perception. The device includes a robot, a 3D vision system, and an industrial control host computer. The robot has a six-degree-of-freedom industrial robotic arm and an end effector. The 3D vision system is installed at the end of the robotic arm or above the workspace to acquire three-dimensional scene point cloud data of the workspace. The industrial control host computer is configured to run vision processing algorithms, occlusion inference models, and motion planning controllers.

[0030] Another embodiment of the present invention provides a robot push-grab collaborative operation method based on active interactive perception. The robot push-grab collaborative operation device based on active interactive perception of the present invention uses this method to complete the pushing of obstacles in the workspace and the grasping of workpieces to be grasped. Specifically, this embodiment uses a box of randomly stacked metal automotive steering knuckles (the steering knuckle surfaces are oily, reflective, and have complex geometric protrusions) as the application scenario, and steering knuckle A in the box as the workpiece to be grasped. The robot push-grab collaborative operation method based on active interactive perception of the present invention is described in detail below. (Refer to...) Figure 1 The method specifically includes the following steps: Step 1: Environmental understanding and accessibility assessment.

[0031] Step 1.1: Construct a 3D scene point cloud in the workspace and match it with the pre-stored workpiece model. Inversely infer the position and orientation of the workpiece to be grasped in the workspace. Call the preset grasping fingerprint database to perform feature twin mapping and determine the theoretical graspable parts of the workpiece to be grasped.

[0032] Specifically, in this embodiment, a 3D vision system is used to acquire depth point cloud data and RGB color images of all the randomly stacked metal car steering knuckles in the bin. The visual coordinate system is transformed to the robot base coordinate system through a hand-eye calibration matrix, thereby constructing a high-precision 3D scene point cloud. When the system uses a point cloud registration algorithm to match the 3D scene point cloud data with the pre-stored CAD model of the workpiece, the 3D vision system can capture at least one local feature of the workpiece to be grasped before reverse inference can be performed. Even if only some features of the workpiece to be grasped are exposed in the field of view, the system can still reverse infer its position and orientation (complete 6D pose) inside the stack based on rigid body geometric constraints. Furthermore, if the constructed 3D scene point cloud fails to match the effective target features of the workpiece to be grasped after matching it with the pre-stored workpiece model, the zero-target fusion strategy is executed. That is, at this time, all the workpieces to be grasped are in a fully occluded state, and the system determines that the current scene cannot be grasped. In this case, the subsequent steps are skipped and the global blind state perturbation strategy is directly triggered so that the robotic arm can significantly flip the workpieces or vibrate the material box in the workspace until the 3D vision system captures at least one local feature of the workpiece to be grasped.

[0033] The pre-defined grasping fingerprint database refers to the set of key geometric surfaces in the CAD model of the workpiece that are predefined and adapted to the stable gripping of the robot's end effector. The method for feature twin mapping based on the pre-defined grasping fingerprint database is as follows: based on the inversely inferred position and orientation of the workpiece to be grasped in the workspace, the grasping fingerprint region of the workpiece to be grasped in the grasping fingerprint database is virtually reprojected onto the current 3D scene point cloud to determine the spatial coordinates of the theoretically graspable part of the workpiece in the current state.

[0034] In this embodiment, the system identifies a steering knuckle A located in the middle layer of the hopper. At this time, only one connecting arm of steering knuckle A is exposed, while its main body is pressed down by the upper steering knuckle B and partially obscured by steering knuckle C. Although most of steering knuckle A is not visible, the point cloud registration algorithm completes the full pose of steering knuckle A in the virtual environment based on the curvature characteristics of the exposed connecting arm and its corresponding CAD model.

[0035] Step 1.2: The theoretically graspable parts are evaluated using both visual and physical filters to confirm whether the workpiece to be grasped can be directly grasped.

[0036] Specifically, the evaluation using a visual filter refers to calculating the exposure α of the grasping fingerprint of the workpiece to be grasped. The specific method is as follows: using the depth comparison method based on camera viewpoint projection, the grasping fingerprint surface after feature twin mapping is discretized into a uniform set of sampling points. The effective proportion of the grasping fingerprint region of the workpiece to be grasped after virtual reprojection is calculated that is not occluded by environmental obstacles. The criterion for determining occlusion by environmental obstacles is: using the depth comparison method based on camera viewpoint projection, along the projection ray from the center of the camera viewpoint to the target sampling point, if a point is detected in the environmental point cloud that is closer to the camera (i.e., has a smaller depth value) and is not within the tolerance range of the target model, then the target sampling point is determined to be occluded.

[0037] The formula for calculating exposure α is α = (S total -S occluded ) / S total ×100%, where S total S represents the theoretical total number of sampling points for fingerprint capture in physical space via feature twin mapping. occluded This represents the number of sampling points within the area that are occluded by environmental obstacles. Evaluation using a physical filter refers to virtual physical collision envelope detection. Specifically, a virtual collision envelope box of the robot's end effector is loaded along the grasping probing path, and the system detects whether the end effector entity interferes with surrounding environmental obstacles during the probing process. Based on the dual evaluation results of the visual and physical filters, the workpiece to be grasped is determined to be directly graspable only when the exposure α is greater than a preset threshold (60%) and the virtual physical collision envelope detection result is no interference. Otherwise, the workpiece is determined not to be directly graspable, and an active interaction strategy is executed to push away obstacles obscuring the workpiece. The active interaction strategy includes generating the optimal pushing path (step 2), obstacle pushing (step 3), and steady-state grasping (step 4).

[0038] In this embodiment, the visual filter evaluation found that the exposure α of steering knuckle A was only 30%, which is less than the preset threshold; and the physical filter evaluation found that if steering knuckle A is forcibly grasped, the left side of the end effector will have a rigid collision with the steering knuckle B above it. Therefore, it is determined that steering knuckle A cannot be directly grasped, and an active interaction strategy needs to be implemented.

[0039] Step 2: Generate the optimal migration path.

[0040] Step 2.1: Construct a local causal topology graph of the target and obstacles to determine the set of obstacles.

[0041] Specifically, the method for constructing the target-obstacle local causal topology graph is as follows: taking the occluded or spatially restricted workpiece to be grasped as the root node, and taking the envelope of the grasping fingerprint tool path that occupies the workpiece in space and / or the set of environmental point clouds that directly cover the model above the workpiece as interference nodes, and establishing directed edges from the upper object to the lower object based on the physical support relationship, thereby obtaining the target-obstacle local causal topology graph; the obstacle set includes all interference nodes in the local causal topology graph.

[0042] In this embodiment, based on the analysis of the local causal topology graph, it is determined that: steering knuckle A is the root node, and steering knuckle B is the interference node that prevents steering knuckle A from being grasped (i.e., steering knuckle B is a critical obstacle, while steering knuckle C is an invalid obstacle that does not affect the grasping of steering knuckle A and does not need to be moved). As long as steering knuckle B is removed, steering knuckle A can be exposed. In the local causal topology graph of this embodiment, the directed edge from the upper object to the lower object is the directed edge from steering knuckle B to steering knuckle A.

[0043] Step 2.2: Traverse the discrete moving directions in the plane for each obstacle in the obstacle set and calculate the overall efficiency cost ratio (Cost) of moving each obstacle along the discrete moving direction. Determine the optimal moving path based on the moving direction corresponding to the minimum overall efficiency cost ratio (Cost).

[0044] Specifically, based on the results of traversing the discrete movement directions of each obstacle in the obstacle set along its plane, it is necessary to calculate the overall efficiency cost ratio (Cost) for moving each obstacle along each discrete movement direction. The overall efficiency cost ratio (Cost) is calculated using the following formula: Cost = (w1·V cloud +w2·D clutter ) / (G cavity ), where V cloud V is the resistance term. cloud V is the estimated physical resistance based on the bounding box volume or projected area of ​​the obstacle point cloud. It represents the estimated physical resistance that needs to be overcome when performing a pushing operation on a specific obstacle. cloudThe value range of D is 0-1; clutter For the environmental item, D clutter It is a quantitative assessment of the spatial margin ahead based on 3D scene point clouds, representing the degree of environmental congestion in the direction of movement, D. clutter The value range of G is 0-1; cavity For the gain term, G cavity G refers to the effective gain of the gripping cavity of the workpiece to be gripped. The gripping cavity refers to the physical space required by the robot's end effector to perform the gripping action. cavity The value range of w1 is 0-1; w1 and w2 are both weighting coefficients, w1+w2=1 and the value range of w1 and w2 is 0-1. When the workpiece in the workspace is a heavy, large workpiece, w1>w2; when the workpiece in the workspace is a light, small workpiece, w1<w2.

[0045] Furthermore, in the scenario of randomly stacked objects, assuming that the material density distribution of workpieces in the same batch is consistent, the mass of an obstacle is positively correlated with its point cloud volume or bottom projection area. The system can obtain the resistance term V by calculating the geometric features of the obstacle point cloud on the pushing support surface to equivalently evaluate the frictional resistance. cloud V cloud The calculation formula is: V cloud =(k·S proj ) / F max_safe , of which S proj S is the projected area of ​​the smallest bounding box of the obstacle point cloud on the moving plane. proj S is positively correlated with frictional resistance. proj The larger the value, the greater the volume and mass of the object, and the higher the frictional resistance that may be generated at the contact surface; k is a preset resistance mapping coefficient, with a value ranging from 0.1 to 10. k is an empirical constant that comprehensively reflects the density properties and surface friction characteristics of the workpiece material; F max_safe V is the preset maximum safe thrust threshold that the robot is allowed to output in this workspace; cloud The value of V strictly falls within the interval 0-1. cloud The closer V is to 0, the smaller and lighter the obstacle, and the less effort (lower cost) it takes to move it. cloud When the value is greater than 1, the system determines that the pushing operation carries the risk of overload shutdown or serious damage to the workpiece, and sets the overall efficiency cost ratio (Cost) corresponding to the pushing direction to infinity and discards it.

[0046] D clutter The following method is used to obtain the virtual displacement envelope corridor constructed by the system along the candidate displacement direction, where the environmental congestion level D is... clutterDefined as the proportion of space within the shifting envelope corridor that is occupied by other environmental obstacles, i.e., D clutter =V obs / V corridor , where V corridor V represents the theoretical total volume of the shifting envelope corridor constructed by the system. obs V represents the total volume of other environmental obstacles located within the pushing envelope corridor, detected by visual point cloud (indicating that this space must be able to accommodate the obstacle to be pushed); the pushing envelope corridor refers to the volume of the space swept by the obstacle to be pushed plus a preset safety margin. obs This can be equivalent to the number of obstacle point clouds or voxel grids within the push envelope corridor. When the candidate push direction is completely open in front, D... clutter =0 indicates no collision obstacles; when D clutter When the value approaches 1, it indicates that the space in front is severely occupied (dead zone). If the system chooses the current direction of movement, it will be subject to a high penalty, thereby guiding the system to choose other, more open directions of movement.

[0047] G cavity To achieve effective gain in grasping the workpiece's gripping cavity, i.e., after the system pre-simulates moving this obstacle, can a sufficiently large gap be created around the workpiece to be gripped to accommodate the end effector and complete the action of gripping the workpiece? Specifically, G cavity The gain term G is obtained by loading the virtual motion envelope box required for the robot's end effector to complete the closing action into the 3D scene point cloud using the following method. cavity G is defined as the proportion of the physical interference volume within the virtual motion envelope that can be eliminated by the pushing operation. cavity =(V t0 -V t1 ) / V t0 , where V t0 The initial interference volume is the total volume of the virtual motion envelope box that the current obstacle point cloud intrudes into and occupies before the push operation; V t1 To predict the residual interference volume, which is the volume of the obstacle point cloud that remains within the virtual motion envelope after the system virtually simulates the obstacle in a specific direction and distance. G cavity The larger the value of G, the more effectively the obstacle can be moved away in this direction; when G cavity When =1, it means that pushing along this direction can completely clear the interference and release 100% of the collision-free space that satisfies the lower claw condition. The system will preferentially select this type of high-gain pushing vector.

[0048] Furthermore, the migration path includes the migration direction and the migration distance along that direction. The method for determining the optimal migration path based on the migration direction corresponding to the minimum overall efficiency cost is as follows: After calculating the overall efficiency cost (Cost) of migrating each obstacle along each discrete migration direction, obviously infeasible solutions are first eliminated. For example, if the environmental term D in a certain direction... clutter >0.9 (almost a complete blind spot ahead), or resistance term V cloud >1 (i.e., exceeding the maximum safe thrust threshold of the robotic arm); then compare the remaining effective Cost values, select the pushing direction corresponding to the lowest overall efficiency cost ratio as the optimal pushing direction, and then perform virtual simulation of the obstacle along the selected optimal pushing direction to obtain the pushing distance along the optimal pushing direction. During the virtual simulation, when V t1 The minimum translation amount required to reduce the obstacle to the point where it no longer obstructs the robot's expected grasping and feeding path is the distance to be pushed along the optimal pushing direction. Thus, the optimal pushing path for pushing away the obstacle is determined.

[0049] In determining the optimal migration path, a threshold value of overall efficiency cost is preset. max (Cost) max (Set to 1.5), if the minimum overall performance cost ratio (Cost) is greater than Cost. max If the global blind state perturbation strategy is triggered (same as the global blind state perturbation strategy in step 1.1), the system state machine is forcibly reset and returns to the environmental understanding and reachability assessment step (step 1).

[0050] In this embodiment, the system simulates three directions of pushing the steering knuckle B: Direction 1 (push to the left): will hit the wall of the hopper (Environmental item D) clutter (Extremely high).

[0051] Specifically, the 3D vision system detected that most of the space within the leftward-moving envelope corridor was occupied by the sidewall of the material bin, and further calculated the environmental term D. clutter It is 0.95. Because D clutter A value greater than 0.9 indicates that the area in front of the direction of movement is almost entirely a dead zone. The system will directly set the Cost corresponding to that direction of movement to infinity and discard it.

[0052] Direction 2 (pushing to the right): There is open space in the direction of movement, but the resistance is estimated to be large (resistance term V). cloud high).

[0053] Specifically, there is open space ahead in the direction of movement, and the crowding is relatively low. Calculations show that D... clutterThe value is 0.2; however, due to the large area of ​​the interconnected workpieces in this direction, the equivalent physical resistance is extremely high, and the calculated value of V is... cloud It is 0.85; further calculations yield G cavity The value is 0.5. Therefore, the overall efficiency cost ratio corresponding to direction 2 is Cost = (0.5 × 0.85 + 0.5 × 0.2) / 0.5 = 1.05, which is denoted as Cost2.

[0054] Direction 3 (pushing backward): This pushing direction is smooth, and after pushing open, it can create a large space for the lower claw of steering knuckle A (G). cavity maximum).

[0055] Specifically, the direction of movement is unobstructed, and calculations show that D... clutter The value is 0.1; the contact area is small in this direction of movement, and the estimated resistance term V is... cloud It is at a low to medium level, and V is calculated to be... cloud The value is 0.4; and pushing it backward can almost completely expose the effective gripping surface of the lower steering knuckle A, resulting in a huge spatial gain. Calculations show that G... cavity The value is 0.95. Therefore, the overall efficiency cost ratio corresponding to direction 3 is Cost = (0.5 × 0.4 + 0.5 × 0.1) / 0.95 = 0.26, which is denoted as Cost3.

[0056] After global optimization and comparison, Cost3 < Cost2. With direction 3 as the optimal pushing direction, the system performs virtual simulation along the selected direction 3. The calculation shows that when the steering knuckle B moves backward by 10cm, it will no longer block the expected grasping path of the robotic arm to the bottom steering knuckle A.

[0057] Step 3: Push away the obstacle.

[0058] The robot moves along the optimal moving path determined in step 2. When the distance between the robot's end effector and the target obstacle is 5-10mm, it switches to variable impedance control mode to move the target obstacle and performs physical deadlock detection during the moving process.

[0059] Specifically, the physical deadlock detection method is as follows: a preset pushing resistance threshold F is used. safe During the migration process, the migration resistance F is continuously monitored. ext The dynamic relationship between the actual moving speed v of the end effector and the monitored pushing resistance F ext The surge exceeded the drag resistance threshold F. safe If the actual moving speed v of the end effector suddenly drops to 0 and remains so for 0.1-0.5s, it is determined that mechanical coupling or static friction lock-up has occurred.

[0060] When a physical deadlock is detected, a high-frequency perturbation strategy is triggered: linear feed is paused, and high-frequency, low-amplitude sinusoidal vibrations are superimposed on the tangential or normal directions of the contact surface. Alternating shear force or dynamic normal pressure changes are used to decouple the contact state from the static friction boundary. After employing the high-frequency perturbation strategy, the resistance F is pushed away. ext Recover to less than the pushing resistance threshold F safe If the actual moving speed v of the end effector recovers, it is determined to be unlocked, and the robot resumes pushing and continues to push. If the high-frequency perturbation strategy cannot unlock, the corresponding obstacle is marked as a physical deadlock node in the local causal topology graph and the process returns to step 2 to replan the suboptimal pushing path for pushing again.

[0061] In this embodiment, the robotic arm switches to variable impedance control mode. As it approaches steering knuckle B, the wrist becomes "soft" to cushion the impact. The robotic arm pushes steering knuckle B along the planned direction 3. When the push reaches 5cm, the bottom of steering knuckle B bulges and unexpectedly hooks onto the steering knuckle below. At this point, the force sensor on the end effector detects the pushing resistance F. ext The surge to 80N (exceeding F) safe Simultaneously, the actual movement speed v of the end effector suddenly dropped to 0, indicating a physical deadlock. At this point, the robotic arm did not forcefully push the steering knuckle B, but immediately triggered a high-frequency perturbation strategy, causing the end effector to perform sinusoidal jitter with a frequency of 15Hz and an amplitude of 2mm in the horizontal direction. Under the effect of jitter, static friction was disrupted, and the coupling loosened. Subsequently, the pushing resistance F of the robotic arm was monitored. ext It descended and continued to smoothly complete the remaining descent work.

[0062] Step 4: Steady-state grasping.

[0063] Step 4.1: The robot is vertically lifted and then reset at the end of the optimal pushing path (first vertically lifted, then horizontally moved to directly above the starting point and then lowered).

[0064] Step 4.2: Reconstruct the 3D scene point cloud in the workspace to confirm whether the workpiece to be grasped can be directly grabbed after the obstacle is moved. If the obstacle blocking the workpiece to be grasped has been removed and both the visual filter and the physical filter have passed the evaluation, then proceed to step 4.3 to directly grasp the workpiece to be grasped; otherwise, with the current scene as the initial state, reconstruct the local causal topology graph of the target and obstacle, and continue with the subsequent steps.

[0065] In this embodiment, a 3D vision system is used to re-photograph the material box globally and reconstruct the three-dimensional scene point cloud. After system comparison, it is confirmed that the steering knuckle B that caused the occlusion has been effectively removed, the exposure α of the grasping fingerprint of the workpiece steering knuckle A below has reached 85% (greater than the preset threshold), and sufficient lower claw space has been released around the steering knuckle A. That is, the virtual physical collision envelope detection has also passed, and the steering knuckle A can be grasped directly.

[0066] Step 4.3: Grab the workpiece to be grabbed.

[0067] Specifically, the method for grasping the workpiece is as follows: the robot moves from its reset starting point to directly above the workpiece and then vertically downwards to perform the grasping action. The moment the robot's end effector contacts the surface of the workpiece, the controller rapidly reduces the normal contact force in the Z-axis direction to zero, achieving a "light touch" to avoid unexpected displacement or surface scratches caused by rigid downward pressure. During the downward movement and closing of the robot's end effector, the system maintains moderately compliant control on the horizontal plane, allowing the gripper to naturally conform to and fit the surface contour of the workpiece under the closing force of two fingers, preventing extreme rigidity. The lateral squeezing force generated by the closure action pushes the workpiece to be grasped away. Compliant control means that the robot's arm maintains a moderately low impedance characteristic on the horizontal plane. The low impedance characteristic is achieved by dynamically adjusting the stiffness parameter in the Cartesian space impedance controller: the translational stiffness of the robot's arm in the X and Y axes is set to 50-500 N / m; after the end effector confirms that the grasping closure is successful, the control mode of the robot arm is switched to stiffness locking mode to ensure that the workpiece to be grasped can remain stable under subsequent heavy load lifting conditions, and then the workpiece to be grasped is removed from the material box, completing the push-grab coordinated operation of the workpiece to be grasped.

[0068] This invention addresses the challenges of workspaces with scattered and stacked workpieces by proposing a robot-assisted push-grab collaborative operation method based on proactive interactive perception. By filtering out key obstacles that obstruct the workpiece to be grasped and proactively planning the optimal path for moving these obstacles to achieve interaction with them, the obstruction is eliminated, enabling precise grasping of the workpiece. Combined with dual evaluation of visual and physical filters and a strategy-based circuit breaker mechanism, the risk of collision damage to the workpiece can be reduced at its source.

[0069] The above embodiments of the present invention are only for illustrating the technical concept and features of the present invention, and are intended to enable those skilled in the art to understand the content of the present invention and implement it accordingly. They should not be construed as limiting the scope of protection of the present invention. All equivalent changes or modifications made in accordance with the spirit and essence of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A robot push-grab collaborative operation method based on active interactive perception, characterized in that, Includes the following steps: Environmental understanding and accessibility assessment: Construct a 3D scene point cloud within the workspace and match it with a pre-stored workpiece model. Inversely infer the position and orientation of the workpiece to be grasped in the workspace. Call a preset grasping fingerprint database to perform feature twin mapping to determine the theoretically graspable parts of the workpiece to be grasped. The theoretically graspable parts are then evaluated by both visual and physical filters to confirm whether the workpiece to be grasped can be directly grasped. When the workpiece to be grasped cannot be directly grasped, an active interaction strategy is executed. The active interaction strategy includes generating the optimal pushing path, obstacle pushing, and steady-state grasping. Generate the optimal migration path: Construct a local causal topology graph of the target and obstacles to determine the set of obstacles. For each obstacle in the set of obstacles, traverse its discrete migration direction in the plane and calculate the overall efficiency cost ratio (Cost) of migrating each obstacle along the discrete migration direction. Determine the optimal migration path based on the migration direction corresponding to the minimum overall efficiency cost ratio (Cost). Obstacle removal: Move obstacles along the optimal path and perform physical deadlock detection during the process. When a physical deadlock is detected, a high-frequency perturbation strategy is triggered. If the high-frequency perturbation strategy can unlock the obstacle, the removal continues. If the high-frequency perturbation strategy cannot unlock the obstacle, the corresponding obstacle is marked as a physical deadlock node in the local causal topology graph and a suboptimal removal path is replanned for removal. Steady-state grasping: After the robot is vertically lifted at the end of the pushing path, it resets and reconstructs the 3D scene point cloud in the workspace to confirm whether it can directly grasp the workpiece after pushing the obstacle. If the obstacle blocking the workpiece has been removed and both the visual filter and the physical filter have passed the evaluation, the workpiece is grasped directly. Otherwise, the local causal topology graph of the target and obstacle is reconstructed with the current scene as the initial state.

2. The robot push-grasp collaborative operation method based on active interactive perception according to claim 1, characterized in that, The condition for reverse inferring the position and orientation of the workpiece to be grasped in the workspace after matching the 3D scene point cloud in the workspace with the pre-stored workpiece model is that the vision system can capture at least one local feature of the workpiece to be grasped.

3. The robot push-grasp collaborative operation method based on active interactive perception according to claim 2, characterized in that, The preset fingerprint database refers to the set of key geometric surfaces predefined in the CAD model of the workpiece, adapted to the stable gripping of the robot's end effector.

4. The robot push-grasp collaborative operation method based on active interactive perception according to claim 3, characterized in that, The feature twin mapping method is as follows: based on the position and orientation of the workpiece to be grasped in the workspace inverse inference, the grasping fingerprint area of ​​the workpiece to be grasped in the grasping fingerprint database is virtually reprojected onto the current three-dimensional scene point cloud to determine the spatial coordinates of the theoretically graspable part of the workpiece to be grasped in the current state.

5. The robot push-grasp collaborative operation method based on active interactive perception according to claim 4, characterized in that, Evaluation using the aforementioned visual filter refers to calculating the exposure α of the grasping fingerprint of the workpiece to be grasped. Specifically, the method involves using a depth comparison method based on camera viewpoint projection to discretize the grasping fingerprint surface after feature twinning mapping into a uniform set of sampling points. The effective proportion of the grasping fingerprint region of the workpiece to be grasped after virtual reprojection, which is not obscured by environmental obstacles, is then calculated. The formula for calculating the exposure α is α = (S... total -S occluded ) / S total ×100%, where S total S represents the theoretical total number of sampling points for fingerprint capture in physical space via feature twin mapping. occluded This represents the number of sampling points in the area that are obscured by environmental obstacles. The evaluation using the physical filter refers to virtual physical collision envelope detection. The specific method is as follows: load the virtual collision envelope box of the robot's end effector along the grasping and probing path, and detect whether the end effector entity will interfere with the surrounding environmental obstacles during the probing process.

6. The robot push-grasp collaborative operation method based on active interactive perception according to claim 5, characterized in that, When the exposure α is greater than a preset threshold and the virtual physical collision envelope detection result is no interference, it is determined that the workpiece to be grabbed can be directly grabbed; otherwise, an active interaction strategy is executed.

7. The robot push-grasp collaborative operation method based on active interactive perception according to claim 3, characterized in that, In the environmental understanding and accessibility assessment steps, if the constructed 3D scene point cloud fails to match the effective target features of the workpiece to be grasped after matching it with the pre-stored workpiece model, then all the workpieces to be grasped are in a state of full occlusion. The system determines that the current scene cannot be grasped and directly triggers a global blind state perturbation strategy to make the robot flip the workpieces or vibrate the bins in the workspace until the vision system captures at least one local feature of the workpiece to be grasped.

8. The robot push-grasp collaborative operation method based on active interactive perception according to claim 7, characterized in that, The method for constructing a local causal topology graph of the target and obstacles is as follows: taking the occluded or spatially restricted workpiece to be grasped as the root node, and taking the envelope of the grasping fingerprint tool path that occupies the workpiece in space and / or the set of environmental point clouds that directly cover the model of the workpiece to be grasped as interference nodes, and establishing directed edges from the upper object to the lower object based on the physical support relationship, thereby obtaining the local causal topology graph of the target and obstacles; the set of obstacles includes all interference nodes in the local causal topology graph.

9. The robot push-grasp collaborative operation method based on active interactive perception according to claim 8, characterized in that, In the step of generating the optimal migration path, based on the results of traversing the discrete migration directions of each obstacle in the obstacle set along its plane, it is necessary to calculate the overall efficiency cost ratio (Cost) for migrating each obstacle along each discrete migration direction. The overall efficiency cost ratio (Cost) is calculated using the following formula: Cost = (w1·V cloud +w2·D clutter ) / (G cavity ), where V cloud V is the resistance term. cloud V is the estimated physical resistance based on the bounding box volume or projected area of ​​the obstacle point cloud. It represents the estimated physical resistance that needs to be overcome when performing a pushing operation on a specific obstacle. cloud The value range of D is 0-1; clutter For the environmental item, D clutter It is a quantitative assessment of the spatial margin ahead based on 3D scene point clouds, representing the degree of environmental congestion in the direction of movement, D. clutter The value range of G is 0-1; cavity For the gain term, G cavity This refers to the effective gain of the gripping cavity of the workpiece to be gripped, where the gripping cavity refers to the physical space required by the robot's end effector to perform the gripping action. cavity The value range of w1 is 0-1; w1 and w2 are both weighting coefficients, w1+w2=1 and the value range of w1 and w2 is 0-1. When the workpiece in the workspace is a heavy, large workpiece, w1>w2; when the workpiece in the workspace is a light, small workpiece, w1<w2.

10. The robot push-grasp collaborative operation method based on active interactive perception according to claim 9, characterized in that, The V cloud The V is obtained by calculating the geometric features of the obstacle point cloud on the pushing support surface to equivalently evaluate the frictional resistance. cloud V is calculated using the following formula: cloud =(k·S proj ) / F max_safe , of which S proj S is the projected area of ​​the smallest bounding box of the obstacle point cloud on the moving plane. proj It is positively correlated with frictional resistance; k is a preset resistance mapping coefficient, and the value of k ranges from 0.1 to 10; F max_safe This is the preset maximum safe thrust threshold that the robot is allowed to output in this workspace; when V cloud When the value is greater than 1, the system determines that the pushing operation carries the risk of overload shutdown or serious damage to the workpiece, and sets the overall efficiency cost ratio (Cost) corresponding to the pushing direction to infinity and discards it.

11. The robot push-grasp collaborative operation method based on active interactive perception according to claim 9, characterized in that, The D clutter The following method is used to obtain the virtual displacement envelope corridor constructed by the system along the candidate displacement direction, where the environmental congestion level D is... clutter Defined as the proportion of space within the shifting envelope corridor that is occupied by other environmental obstacles, i.e., D clutter =V obs / V corridor , where V corridor V represents the theoretical total volume of the shifting envelope corridor constructed by the system. obs V represents the total volume of other environmental obstacles located within the shifting envelope corridor, detected by visual point cloud; the shifting envelope corridor refers to the spatial volume swept across by the obstacle to be shifted plus a preset safety margin. obs This can be equivalent to the number of obstacle point clouds or voxel grids within the shifting envelope corridor; if D clutter When the value is greater than 0.9, it indicates that there is a dead angle ahead in the direction of movement. The overall performance cost ratio (Cost) corresponding to the direction of movement is set to infinity and eliminated.

12. The robot push-grasp collaborative operation method based on active interactive perception according to claim 9, characterized in that, The G cavity The gain term G is obtained by loading the virtual motion envelope box required for the robot's end effector to complete the closing action into the 3D scene point cloud using the following method. cavity G is defined as the proportion of the physical interference volume within the virtual motion envelope that can be eliminated by the pushing operation. cavity =(V t0 -V t1 ) / V t0 , where V t0 The initial interference volume is the total volume of the virtual motion envelope box that the current obstacle point cloud intrudes into and occupies before the push operation; V t1 To predict the residual interference volume, which is the volume of the obstacle point cloud that remains within the virtual motion envelope after the system virtually simulates the obstacle in a specific direction and distance.

13. The robot push-grasp collaborative operation method based on active interactive perception according to claim 12, characterized in that, The method for determining the optimal migration path based on the migration direction corresponding to the minimum overall efficiency cost ratio (Cost) is as follows: After calculating and comparing the overall efficiency cost ratio (Cost) of each obstacle along each discrete migration direction, the migration direction corresponding to the minimum overall efficiency cost ratio (Cost) is selected as the optimal migration direction. Then, a virtual simulation is performed on the obstacles along the optimal migration direction to obtain the migration distance along that direction. During the virtual simulation, when V... t1 The minimum translation amount corresponding to reducing the obstacle to no longer obstruct the robot's expected grasping and feeding path is the pushing distance along the optimal pushing direction; Preset overall performance cost threshold (Cost) max If the minimum overall performance cost is greater than the cost max This directly triggers the global blind state perturbation strategy and forces the system state machine to jump and reset, returning to the environmental understanding and reachability assessment steps.

14. The robot push-grasp collaborative operation method based on active interactive perception according to claim 1, characterized in that, In the obstacle pushing step, when the robot starts the pushing operation and the distance between the robot end and the target obstacle is 5-10mm, it switches to the variable impedance control mode to push the target obstacle.

15. The robot push-grasp collaborative operation method based on active interactive perception according to claim 3, characterized in that, The physical deadlock detection method is as follows: a preset pushing resistance threshold F is used. safe During the migration process, the migration resistance F is continuously monitored. ext The dynamic relationship between the actual moving speed v of the end effector and the monitored pushing resistance F ext The surge exceeded the drag resistance threshold F. safe If the actual moving speed v of the end effector suddenly drops to 0 and remains so for 0.1-0.5s, it is determined that mechanical coupling or static friction lock-up has occurred.

16. The robot push-grasp collaborative operation method based on active interactive perception according to claim 15, characterized in that, The high-frequency perturbation strategy involves pausing linear feed and superimposing high-frequency, low-amplitude sinusoidal vibrations in the tangential or normal directions of the contact surface; when the high-frequency perturbation strategy is adopted, the thrust F... ext Recover to less than the pushing resistance threshold F safe If the actual moving speed v of the end effector recovers, it is determined to be unlocked, and the robot resumes pushing and continues to complete the pushing.

17. The robot push-grasp collaborative operation method based on active interactive perception according to claim 3, characterized in that, In the steady-state gripping step, the method for gripping the workpiece is as follows: the robot moves from the reset starting point to directly above the workpiece and vertically downwards to perform the gripping. At the instant the robot's end effector contacts the surface of the workpiece, the controller rapidly attenuates the normal contact force in the Z-axis direction to 0. During the downward movement and closing of the robot's end effector, the robot's arm maintains low impedance characteristics on the horizontal plane. The low impedance characteristics are achieved by dynamically adjusting the stiffness parameters in the Cartesian space impedance controller. The translational stiffness of the robot's arm in both the X and Y axes is set to 50-500 N / m. After the end effector confirms successful gripping closure, a stiffness locking mode is adopted to move the workpiece out of the workspace, completing the push-grip collaborative operation of the workpiece.

18. A robot push-grab collaborative operation device based on active interactive perception, characterized in that, The robot push-grab collaborative operation method based on active interactive perception as described in any one of claims 1-17 is used to complete the pushing of obstacles in the workspace and the grasping of workpieces to be grasped.

Images