Robot pushing and grabbing cooperation method based on two-way independent competitive network

Abstract

The invention provides a robot pushing and grabbing cooperation method based on a two-way independent competitive network. Firstly, an original image of a scene is obtained, and an orthogonal color height map and a depth height map are generated; a double-channel independent competition network in a full convolution form is constructed, and Q value distribution information of pushing and grabbing actions at a pixel level is predicted in parallel by using the decoupled state value flow and action advantage flow; a cross-task independent competition decision-making mechanism is established, through an active exploration and hierarchical decision-making method and in combination with an optimization strategy, a mechanical arm is driven to execute physical interaction, and meanwhile exploration and pushing and grabbing cooperation tasks are completed; designing a joint reward function and a multi-task loss function to calculate model loss, and improving model performance based on a hierarchical training mechanism; compared with the prior art, the motion value evaluation is accurate, the invalid exploration rate is low, the cooperation efficiency is high, closely stacked objects can be separated autonomously through the pushing motion, and the method has extremely high operation success rate and robustness in a complex unstructured scene.

Description

Technical Field

[0001] This invention belongs to the field of robot intelligent control and computer vision, specifically relating to a robot push-grab cooperative decision-making method based on a dual-path independent competitive network. Background Technology

[0002] With the acceleration of intelligent manufacturing and Industry 4.0, the operational logic of industrial robots is shifting from traditional pre-programmed, structured environments to unstructured, complex, and dynamic scenarios. In these emerging scenarios, robotic arms must possess the ability to make autonomous decisions and complete efficient interactions in chaotic, stacked environments.

[0003] Robotic grasping, as the most fundamental ability of a robotic arm to interact with its environment, relies on the core principle of calculating a stable and reliable grasping pose in real time based on sensor data. Traditional grasping methods are mainly based on analytical models, depending on the precise 3D geometric CAD model of the target object and rigorously calibrated environmental parameters, using static analysis to determine the closure of the grasping force. However, when faced with chaotic scenarios involving unknown shapes, tightly stacked objects, and severe physical occlusion, traditional methods often fail to meet the stringent requirements of real-time performance and robustness in industrial settings due to difficulties in obtaining accurate geometric priors, extreme sensitivity to sensor noise, and the enormous computational overhead of inverse kinematics solutions.

[0004] In recent years, the rise of deep reinforcement learning has provided a new path for end-to-end visual grasping. By combining visual perception features with action policy mapping, agents can autonomously learn policies from raw images to operational poses through continuous trial and error with the environment. Traditional algorithms typically use single-path fully convolutional Q-networks to directly regress action value. These networks often struggle to effectively distinguish between the inherent value of the environmental state and the incremental advantage brought by specific rotational actions. This coupled evaluation method easily leads to overestimation of Q-values, causing severe oscillations in the training process and resulting in slow policy convergence. In chaotic scenes, the effective operational area usually occupies only a very small proportion of the total workspace. Existing algorithms often exhibit numerous empty pushes and grasps in the early stages of exploration, i.e., performing interactions in ineffective locations such as far from the object or the object's center. Due to the extremely sparse reward signal, a large number of invalid attempts severely consume computational resources and reduce sample utilization. Existing collaborative logic is often too simplistic and lacks specific design for extreme challenging scenarios such as tightly packed objects or zero-gap arrangements. When objects are tightly stacked, robotic arms often cannot directly perform effective grasping. How to accurately create physical space for grasping through pushing actions, and establish a deep game mechanism that promotes grasping through pushing and co-evolution, is a difficult problem that urgently needs to be solved in the current field. Summary of the Invention

[0005] The purpose of this invention is to provide a robot push-grab cooperation method based on a dual-path independent competitive network to solve the problems mentioned in the background art.

[0006] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows:

[0007] A robot push-grasping cooperative method based on a dual-path independent competitive network includes the following steps:

[0008] (1) The acquired original image is preprocessed and a color height map and a depth height map are generated by orthogonal projection transformation;

[0009] (2) Construct a fully convolutional dual-path independent competitive network model, and output the state value, action advantage and aggregation information of the pushing and grasping actions at the pixel level through parallel output of network branches;

[0010] (3) Establish a cross-task independent competitive decision-making mechanism, and drive the robotic arm to perform physical interaction by actively exploring and hierarchical decision-making methods, combined with optimization strategies, while completing exploration and push-grab collaborative tasks.

[0011] (4) Design a joint reward function and a multi-task loss function to calculate the model loss, and improve the model performance based on a hierarchical training mechanism.

[0012] (5) Deploy the trained model on the computing platform to realize real-time reasoning from visual input to action output, and meet the online operation requirements of the robotic arm in complex working environments.

[0013] As a further technical solution of the present invention, the preprocessing process in step (1) is as follows: 1) The original three-dimensional point cloud data is reprojected onto the working space reference plane using the perspective transformation method to generate a color height map and a depth height map with consistent spatial resolution; 2) The generated height map group is rotated at a certain angle step to construct an input state space covering the operation direction, so as to enhance the network's robustness in perceiving the rotation posture of the object.

[0014] As a further technical solution of the present invention, in step (2), a dual-path independent competitive network model in the form of full convolution is constructed. The specific steps are as follows: 1) A feature extraction backbone network with a full convolution structure is adopted, and multi-scale feature reuse is realized through a dense connection mechanism to output a high-dimensional feature tensor that integrates scene color semantics and geometric topology features; 2) The feature tensor is fed into the symmetrical push branch and the grab branch respectively. Each branch is decoupled into a state value flow and an action advantage flow using parallel convolutional layers; wherein the state value flow outputs a single-channel state value map to represent the inherent operational potential of each pixel position in the scene; the action advantage flow outputs a single-channel action advantage map to represent the policy incremental advantage under a specific rotation angle; and a prediction map aligned with the input graph space is generated using an aggregation operator.

[0015] As a further technical solution of the present invention, in step (3), a cross-task independent competitive decision-making mechanism is established. Through active exploration and hierarchical decision-making methods, combined with optimization strategies, the robotic arm is driven to perform physical interaction and complete the exploration and push-grab collaborative tasks. The specific steps are as follows: 1) Generate the final action confidence map through the decision model of the exploration strategy and extract the action index with the maximum response; 2) When the exploration strategy successfully discovers the target, the system switches to the coordination strategy and proposes a state classification-based collaborative mechanism. A lightweight action classifier is introduced, and the geometric congestion of the scene and the historical operation state are used as the decision basis; 3) Optimize the strategy. Based on the post-target relabeling mechanism, if the task is judged to have failed, but the end effector is detected to have successfully grabbed a non-target object in the scene, the relabeling process is triggered to recalculate the reward value of the action and store the original sample and the relabeled sample in the experience playback buffer at the same time.

[0016] As a further technical solution of the present invention, in step (4), a joint reward function and a multi-task loss function are designed to calculate the model loss, and the model performance is improved based on a hierarchical training mechanism, specifically as follows:

[0017] 1) A phased guided reward function, including pre-action intention reward and post-action effect reward; 2) Construct a three-component composite loss function covering adversarial induction loss, execution efficiency constraint, and spatial smoothing regularization loss; 3) A hierarchical training mechanism: In the first stage, we introduced a target condition mechanism for network training; in the second stage, we used the captured network trained in the first stage to obtain scene potential energy through its output action value distribution; in the third stage, we adopted an alternating fine-tuning strategy, which includes two alternating sub-steps. Through multiple rounds of alternating iterations, the state distributions of the captured network and the push network gradually align, forming a closed-loop adaptation.

[0018] Compared with the prior art, the beneficial effects of the present invention are:

[0019] (1) By adopting a fully convolutional decoupling architecture, the state value and action advantage are deeply separated, which effectively solves the overestimation problem in dense action space and significantly improves the accuracy of policy evaluation.

[0020] (2) By establishing a cross-task independent competitive decision-making mechanism, the robotic arm can autonomously coordinate the pushing and grasping action logic according to environmental feedback. In particular, in a zero-gap stacking environment, it can create operational space for subsequent grasping through precise pushing and separating actions.

[0021] (3) The fully convolutional design ensures strict spatial alignment between the visual perception state and the physical execution action, improving the real-time response speed and pose control accuracy of the system. Attached Figure Description

[0022] Figure 1This is a flowchart illustrating the overall process of a robot push-grasp cooperative method based on a dual-path independent competitive network as described in this invention.

[0023] Figure 2 This is a topology diagram of the dual-path independent competitive network model described in this invention. Detailed Implementation

[0024] The specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples. The following examples are for illustrative purposes only and are not intended to limit the scope of the invention.

[0025] Referring to Figure 1, the robot push-grasp cooperative method based on a dual-path independent competitive network according to the present invention includes:

[0026] (1) The acquired original image is preprocessed and a color height map and a depth height map are generated by orthogonal projection transformation.

[0027] Furthermore, the specific steps are as follows:

[0028] 1) First, use an RGB-D camera fixedly mounted above the workbench to acquire real-time color images of the scene. With depth map ;

[0029] 2) The system utilizes the camera intrinsic parameter matrix obtained in advance through hand-eye calibration. With homogeneous pose transformation matrix Each pixel in the original image Combine its depth value to reconstruct a 3D point cloud in the world coordinate system;

[0030] 3) Subsequently, the system executes the orthogonal reprojection operator to project the point cloud along the negative direction of gravity onto the preset workspace plane. Set the workspace size to... Set the resolution of the projection output to 10 ... Pixels. With this configuration, each pixel in the generated color heightmap and depth heightmap corresponds precisely to a pixel in physical space. The columnar region;

[0031] 4) Finally, to impart rotation invariance to the network, the heightmap group is... Perform 16 equal-angle rotations with respect to the step size to construct a cover. The sequence of input tensors for the operation direction.

[0032] (2) Construct a dual-path independent competitive network model in the form of full convolution, and output the state value, action advantage and aggregation information of the driving and grasping actions at the pixel level through parallel output of network branches.

[0033] Furthermore, referring to Figure 2, the specific steps for constructing the dual-path independent competition network model and its internal decoupling logic are as follows:

[0034] 1) Feature Extraction Backbone Design: A fully convolutional DenseNet-121 network is used as the backbone. The input receives multimodal height maps, and multi-scale feature reuse is achieved through skip connections in dense blocks. The backbone network outputs a high-dimensional feature tensor with 2048 channels. This tensor depth integrates the geometric topology and color distribution features of the scene;

[0035] 2) Decoupling competing branch mapping: Mapping the feature tensor... The data is fed into the push branch and the grasp branch, respectively. Within each branch, it is further decoupled into two functional streams: the state value stream uses a convolutional sequence to compress the 2048-dimensional features into a single channel, outputting a state value map. This map is used to evaluate the inherent operational potential of the scene at the current pixel location, characterizing the background score indicating the presence of an object or operational feasibility at that location; the advantage value stream outputs an action advantage map using a symmetrical convolutional structure. This is used to evaluate policy increments at specific rotation angles. Because the input is rotationally enhanced, this stream can accurately capture the relative advantages brought by the object's edge orientation;

[0036] 3) Pixel-level value aggregation and heatmap generation: For coordinates The first Each action utilizes aggregation operators to perform centralized aggregation operations:

[0037]

[0038] in, Represents a state. and The state-value-action-advantage value is calculated by mathematically constraining a deep separation between value and advantage, ultimately outputting a spatially aligned push and grab Q-value prediction heatmap.

[0039] (3) Establish a cross-task independent competitive decision-making mechanism. Through active exploration and hierarchical decision-making methods, combined with optimization strategies, drive the robotic arm to perform physical interaction and complete exploration and push-grab collaborative tasks at the same time.

[0040] The specific implementation process of the decision and mapping algorithm is as follows:

[0041] 1) Active Exploration: The decision-making model for exploration strategies consists of the following three core modules:

[0042] The feasibility measurement module uses an existing push network for generalized reasoning. The confidence map output by the network degenerates into a general physical feasibility heatmap, which defines a safe and effective physical boundary for subsequent exploration actions.

[0043] The geometric prior distribution module designs an edge detection operator based on height difference. For each pixel in the height map, its height gradient with neighboring pixels is calculated. If the height difference exceeds a threshold, the point is determined to be a significant edge. This generates an edge feature map, resulting in the final geometric prior distribution.

[0044] The experience inhibition module constructs a corresponding inhibition function for each historical action position that leads to exploration failure. The final failure likelihood distribution is the cumulative effect of the most recent failure records, forcing the exploration strategy to shift to unexplored areas. Based on the above three modules, the active exploration strategy generates the final action confidence map through fusion.

[0045] 2) Hierarchical Decision-Making: After the exploration strategy successfully discovers the target, the system switches to a coordination strategy, proposing a collaborative mechanism based on state classification. A lightweight action classifier is introduced, using the geometric congestion of the scene and historical operation states as the basis for decision-making.

[0046] Target boundary congestion, relative to target mask Perform morphological dilation to construct a mask for the target boundary region. Based on this, the object occupancy rate within the boundary region is calculated. Let... For heightmap values, if Then it is determined that an object occupies the space. Defined as:

[0047]

[0048] in For pixels The area occupied at that location Represents the pixel threshold;

[0049] To eliminate the influence of the target's own size, we introduce the relative crowding norm, which calculates the number of pixels occupied by the boundary relative to the number of pixels of the target itself:

[0050]

[0051] To enable the system to dynamically adjust, we record the number of consecutive failed crawling attempts. ,along with As the number of fetches increases, the system will tend to lower the confidence level of the fetch and instead execute pushes to change the current deadlock.

[0052] We will combine the above features with the maximum action value of the network output. splicing, and The maximum action value of pushing and grabbing actions are represented respectively, forming the decision state vector:

[0053]

[0054] 3) Optimization strategy: Based on the post-event target relabeling mechanism, assuming at time... The state of the agent is The target mask is Execute actions After transitioning to state The original reward obtained was If the task is deemed to have failed, but the end effector is detected to have successfully captured a non-target object in the scene. This triggers the remarking process:

[0055] First, call the segmentation module to obtain the object. Corresponding pixel-level mask Then construct a new virtual transfer sample. The target mask from the original input... Replace with Based on the new mask And the reward function, recalculate the reward value for the action. Because the actual action did indeed grasp the object. According to the definition of the capture reward function, The original sample will be assigned a high score. and relabeled samples At the same time, it is stored in the experience replay buffer.

[0056] (4) Design a joint reward function and a multi-task loss function to calculate the model loss, and improve the model performance based on a hierarchical training mechanism.

[0057] 1) Joint Reward Function: A phased guided reward function, including pre-action intention reward and post-action effect reward; if the pushing trajectory has geometric intersection with the target mask, the action is considered to have the correct operational intention and is given a corresponding reward; if the boundary occupancy value decreases significantly after the pushing action is executed, it indicates that the debris has been successfully moved out of the target's vicinity, creating a positive spatial gain for grasping. At this time, a high reward is given, and the final pushing reward is the sum of the two.

[0058] The reward design for the grasping action aims to balance the accuracy of the grasping position with the success rate of the final task. We designed reward functions before and after the grasping action. To encourage the network to output geometrically reasonable grasping points, an initial reward is given if the predicted grasping center point falls inside the target mask. The gripper center should be as close as possible to the physical centroid of the object, and a corresponding reward is given based on the relative position of the object's centroid and the gripper. In addition to positional accuracy, the alignment of the grasping angle with the object's geometric principal axis also determines the success or failure of the grasp, and a corresponding reward is given based on the parallelism between the gripper's closing direction and the object's principal axis. The final reward before grasping is the sum of these three factors. The reward function after grasping is designed as a simple binary function, and the final grasping reward function is the sum of the rewards before and after grasping.

[0059] 2) Multi-task loss function: Based on this, a three-component composite loss function is constructed, encompassing adversarial induction loss, execution efficiency constraints, and spatial smoothing regularization loss:

[0060]

[0061] in To combat induced loss, To enforce efficiency constraints, This is the spatial smoothing regularization loss.

[0062] 3) Layered Training Mechanism: In the first stage of network training, we introduced a target condition mechanism, enabling the grasping network to quickly establish its perception of object geometric features using a large number of erroneous grasping samples in the early stages of training, laying a solid action foundation for subsequent strategy learning. In the second stage, using the grasping network trained in the first stage, we obtain the scene potential energy through its output action value distribution. When the scene potential energy is determined to have dropped below a threshold, terminal grasping is triggered; otherwise, the system will continue to reconstruct the scene through the push network until the maximum step limit is reached. Although the first two stages train grasping and pushing capabilities respectively, a severe distribution mismatch problem will be faced during joint inference. To eliminate this distribution bias, the third stage adopts an alternating fine-tuning strategy, which includes two alternating sub-steps: fixing the push network and updating the grasping network. This step allows the grasping network to learn more scenes and learn to recognize and utilize the favorable conditions created by the push network. Fixing the updated parameters of the grasping network and updating the parameters of the push network. This step forces the push network to adjust its strategy based on the feedback from the grasping network, creating conditions for the grasping network. Through multiple rounds of alternating iterations, the state distributions of the grasping network and the push network gradually align, forming a closed-loop adaptation.

[0063] (5) Deploy the trained model on the computing platform to realize real-time reasoning from visual input to action output, and meet the online operation requirements of the robotic arm in complex working environments.

[0064] The weights of the converged model are solidified and ported to an industrial computing platform with edge computing capabilities. Utilizing the efficient parallel inference capabilities of a fully convolutional neural network architecture, a collaborative grasping test of a robotic arm in a complex working environment is completed, demonstrating that the proposed method can achieve efficient and robust target acquisition in various unstructured, high-density, and cluttered environments.

[0065] This invention is based on the National Natural Science Foundation of China (62273085) project "Research on Data-Driven Fault Diagnosis Method for Unmanned Mining Trucks under Unknown Driving Conditions in Mines".

Claims

1. A robot push-grasping cooperative method based on a dual-path independent competitive network, characterized in that, Includes the following steps: (1) Construct a dual-path independent competitive network model, output the state value and action advantage of pushing and grasping actions in parallel through network branches, and output aggregated information through value aggregation algorithm; (2) Establish a cross-task independent competitive decision-making mechanism, and drive the robotic arm to perform physical interaction by actively exploring and hierarchical decision-making methods, combined with optimization strategies, while completing exploration and push-grab collaborative tasks. (3) Design a joint reward function and a multi-task loss function to calculate the model loss, and improve the model performance based on a hierarchical training mechanism.

2. The robot push-grasp cooperative method based on a dual-path independent competitive network according to claim 1, characterized in that, In step (1), the constructed dual-path independent competitive network model outputs the state value and action advantage of pushing and grasping actions in parallel through network branches, and outputs aggregated information through a value aggregation algorithm, including: The feature tensors are fed into the push branch and the grasp branch respectively. Within each branch, they are further decoupled into two functional flows. The state value flow uses a convolutional sequence to compress the features into a single channel and outputs a state value map to evaluate the inherent operational potential of the scene at the current pixel position. The advantage value flow uses a symmetrical convolutional structure to output an action advantage map to evaluate the policy increment under a specific rotation angle. For coordinates The first Each action utilizes aggregation operators to perform centralized aggregation operations: in, Represents a state. and The state-value-action-advantage value is calculated by mathematically constraining a deep separation between value and advantage, ultimately outputting a spatially aligned push and grab Q-value prediction heatmap.

3. The robot push-grasp cooperative method based on a dual-path independent competitive network according to claim 1, characterized in that, In step (2), the active exploration method completes the exploration task, including: The decision-making model for exploration strategies consists of the following three core modules: The feasibility measurement module uses an existing push network for generalized reasoning. The confidence map output by the network degenerates into a general physical feasibility heatmap, which defines a safe and effective physical boundary for subsequent exploration actions. The geometric prior distribution module designs an edge detection operator based on height difference. For each pixel in the height map, it calculates the height gradient between the pixel and its neighboring pixels. If the height difference exceeds a threshold, the pixel is determined to be a significant edge, thereby generating an edge feature map and obtaining the final geometric prior distribution. The experience inhibition module constructs a corresponding inhibition function for each historical action position that leads to exploration failure. The final failure likelihood distribution is the cumulative effect of the most recent failure records, forcing the exploration strategy to shift to unexplored areas. Based on the above three modules, the active exploration strategy generates the final action confidence map by fusion.

4. The robot push-grasp cooperative method based on a dual-path independent competitive network according to claim 3, characterized in that, In step (2), the hierarchical decision-making method completes the push-grab collaboration task, including: Once the exploration strategy successfully identifies the target, the system switches to a coordination strategy, proposing a state-classification-based coordination mechanism. This mechanism introduces a lightweight action classifier, using the geometric congestion of the scene and historical operation states as decision-making criteria. Target boundary congestion, relative to target mask Perform morphological dilation to construct a mask for the target boundary region. Based on this, calculate the object occupancy rate within the boundary area, and assume... For heightmap values, if Then it is determined that an object occupies the space. Defined as: in For pixels The area occupied at that location Represents the pixel threshold; To eliminate the influence of the target's own size, we introduce the relative crowding norm, which calculates the number of pixels occupied by the boundary relative to the number of pixels of the target itself: To enable the system to dynamically adjust, we record the number of consecutive failed crawling attempts. ,along with As the number of fetches increases, the system will tend to lower the confidence level of the fetch and instead execute pushes to change the current deadlock. We will combine the above features with the maximum action value of the network output. splicing, and The maximum action value of pushing and grabbing actions are represented respectively, forming the decision state vector: This allows us to determine the final push and grab decision.

5. The robot push-grasp cooperative method based on a dual-path independent competitive network according to claim 4, characterized in that, In step (2), the optimization strategy includes: Based on the post-event target relabeling mechanism, assuming at time... The state of the agent is The target mask is Execute actions After transitioning to state The original reward obtained was If the task is deemed to have failed, but the end effector is detected to have successfully captured a non-target object in the scene, then... This triggers the remarking process: First, call the segmentation module to obtain the object. Corresponding pixel-level mask Then, a new virtual transfer sample is constructed, which is the target mask in the original input. Replace with Based on the new mask And the reward function, recalculate the reward value for the action. Because the actual action did indeed grasp the object. According to the definition of the capture reward function, The original sample will be given a high score. and relabeled samples At the same time, it is stored in the experience replay buffer.

6. The robot push-grasp cooperative method based on a dual-path independent competitive network according to claim 1, characterized in that, In step (3), a joint reward function and a multi-task loss function are designed to calculate the model loss: A phased guided reward function includes pre-action intention reward and post-action effect reward. If the pushing trajectory has a geometric intersection with the target mask, the action is considered to have the correct operational intention and is given a corresponding reward. If the boundary occupancy value drops significantly after the pushing action is executed, it means that the debris has been successfully moved out of the target's vicinity, creating a positive spatial gain for grabbing. At this time, a high reward is given, and the final pushing reward is the sum of the two. The reward design for the grasping action aims to balance the accuracy of the grasping position with the success rate of the final task. We also designed reward functions before and after the grasping action. In order to encourage the network to output geometrically reasonable grasping points, if the predicted grasping center point falls inside the target mask, an initial reward is given; the center of the gripper should be as close as possible to the physical centroid of the object, and a corresponding reward is given based on the relative position of the object's centroid and the gripper. Besides positional accuracy, the alignment of the gripping angle with the object's geometric principal axis also determines the success or failure of the grip. A reward is given based on the parallelism between the gripper's closing direction and the object's principal axis. The final reward before gripping is the sum of these three factors. The reward function after gripping is designed as a simple binary function, and the final gripping reward function is the sum of the rewards before and after gripping. Based on this, a three-component composite loss function is constructed, encompassing adversarial induction loss, execution efficiency constraints, and spatial smoothing regularization loss: To combat induced loss, To enforce efficiency constraints, This is the spatial smoothing regularization loss.

7. A robot push-grasp cooperative method based on a dual-path independent competitive network according to claim 6, characterized in that, In step (3), based on the hierarchical training mechanism, the model performance is improved by including: In the first stage of network training, we introduced a target condition mechanism, enabling the grasping network to quickly establish its perception of object geometric features using a large number of erroneous grasping samples in the early stages of training. This lays a solid action foundation for subsequent strategy learning. In the second stage, using the grasping network trained in the first stage, we obtain scene potential energy through the action value distribution of its output. When the scene potential energy is determined to have dropped below a threshold, terminal grasping is triggered; otherwise, the system will continue to reconstruct the scene through the push network until the maximum number of steps is reached. Although the first two stages train grasping and pushing capabilities respectively, they face a severe distribution mismatch problem during joint inference. To eliminate this distribution bias, the third stage adopts an alternating fine-tuning strategy, which includes two alternating sub-steps: fixing the push network and updating the grasping network. This step allows the grasping network to learn more scenes and learn to recognize and utilize the favorable conditions created by the push network. Fixing the updated parameters of the grasping network and updating the parameters of the push network. This step forces the push network to adjust its strategy based on the feedback from the grasping network, creating conditions for the grasping network. Through multiple rounds of alternating iterations, the state distributions of the grasping network and the push network gradually align, forming a closed-loop adaptation.

Images