A dual-loop strategy-based grasping and releasing decision method for long sequence objects in cluttered stacking scenes

Abstract

The present invention discloses a method for grasping and releasing objects in a long sequence of cluttered stacking scenes based on a dual-loop strategy, which includes two closed-loop strategies. The inner-loop strategy is responsible for improving the accuracy of object matching through active perception. When the robot grasps the object, the strategy will perceive it from multiple perspectives by rotating the object to increase the confidence of the object matching with the object in the target scene. The outer-loop strategy is responsible for executing the grasping and placing operations, combining the uncertain object matching and grasping ability to predict the grasping action, and deciding whether to place it directly at the target position or temporarily place the object in the buffer according to the matching situation of the current object. The method uses a pre-trained visual basic model for object matching, inner-loop strategy learning and inner-loop termination mechanism. The method is more robust to the influence of perception noise, and considers the optimality of the task level through reinforcement learning closed-loop optimization, thereby achieving better task success rate and efficiency.

Description

Technical Field

[0001] The present invention relates to the field of robot intelligent operation, and in particular to a method for grasping and releasing objects in a long sequence in a cluttered stacking scene based on a double-loop strategy. Background Art

[0002] At present, grasping and releasing long sequences of objects in cluttered stacking scenes is a very important task in the field of robot manipulation and a key technology for service robots. Traditional methods rely on known object models, but it is usually difficult for users to provide such information. In recent years, work has begun to study the long-sequence grasping and releasing tasks of multiple objects with unknown object models. Given only an RGB-D image of the initial scene and an RGB-D image of the target scene, the work explores how to rearrange the multiple objects in the initial scene into the state of the target scene.

[0003] Since there is no object model, some recent works mainly use learning-based perception modules to achieve matching between the initial scene and the target scene to generate scene representation. These methods significantly narrow the perception gap in the task of reordering tasks from known objects to unknown objects. However, the actions of these systems at the planning level are derived through heuristic rules or supervised learning, and less attention is paid to the optimality of the task level. This situation is exacerbated by perceptual noise because these heuristic rules may be based on erroneous perceptual results. In addition, existing systems have difficulty optimizing the overall task completion efficiency when dealing with long task sequences. To optimize task-level performance, reinforcement learning is a useful tool. However, there are challenges in directly learning policies in long sequence tasks through reinforcement learning.

[0004] The prior art has the following technical problems:

[0005] 1) The efficiency of long sequence pick-and-place sequences generated by existing technologies in stacking scenarios is generally low;

[0006] 2) Existing technologies have poor generalization to unknown objects;

[0007] 3) Existing technologies have poor robustness to perceptual noise when the object model is unknown;

[0008] 4) Existing techniques do not focus on optimality at the task level. Summary of the invention

[0009] In order to overcome the shortcomings of the prior art, the purpose of the present invention is to provide a method for grasping and releasing objects in a long sequence in a cluttered stacking scene based on a dual-loop strategy. On the basis of the grasping and releasing closed-loop strategy, the present invention introduces an active perception strategy as an inner-loop strategy to form a dual-loop strategy to predict the grasping, looking, and releasing action sequence.

[0010] Specifically, the method includes two closed-loop strategies. The inner-loop strategy is responsible for improving the accuracy of object matching through active perception. When the robot grasps the object, the strategy will perceive it from multiple perspectives by rotating the object to increase the confidence of the object matching with the object in the target scene. The outer-loop strategy is responsible for performing grasping and placement operations. It combines uncertain object matching and grasping capabilities to predict grasping actions, and decides whether to place the object directly at the target position or temporarily place the object in the buffer according to the current object matching situation. The strategy realizes closed-loop optimization through reinforcement learning with task-level feedback. In order to achieve generalization performance, the method uses a pre-trained visual basic model for object matching, inner-loop strategy learning, and inner-loop termination mechanism. Compared with previous work, this method is more robust to the impact of perceptual noise, and considers the optimality of the task level through reinforcement learning closed-loop optimization, thereby achieving better task success rate and efficiency.

[0011] The present invention is achieved through the following technical solutions:

[0012] The present invention discloses a method for grasping and releasing objects in a long sequence in a cluttered stacking scene based on a dual-loop strategy, comprising:

[0013] Step 1: Given the RGB-D image of the target scene, obtain the RGB-D image of the current scene;

[0014] Step 2: Obtain all object bounding boxes in the current and target scenes through the pre-trained open set object detection algorithm and form object image representation;

[0015] Step 3: Extract visual features of all object image representations through the pre-trained visual basic model;

[0016] Step 4: Based on the RGB-D image of the current scene, the pre-trained grasping algorithm model is used to predict the grasping action set in the scene and encode it into grasping action features;

[0017] Step 5: Based on the obtained visual features and grasping action features of all objects in the current and target scenes, the transformer is used to implement the cross-attention mechanism of the visual and action modalities to generate cross-attention features;

[0018] Step 6: Based on the obtained cross-attention features, the probability values ​​of all grasping actions are generated through the grasping strategy network;

[0019] Step 7: Select a grasping action according to the probability value of the grasping action, execute the grasping action with the highest probability value, and obtain the grasped object;

[0020] Step 8: According to the grasped object, the pre-trained visual basic model is used to match the object with all objects in the target scene to obtain the target object that best matches the grasped object and its matching score;

[0021] Step 9: Determine whether to start the active sensing inner loop strategy based on the obtained matching score. If active sensing is performed, execute steps 10-12; otherwise, directly execute step 13;

[0022] Step 10: According to the captured object obtained in step 7 and the target object that best matches the captured object obtained in step 8, obtain the object bounding boxes obtained in step 2 respectively, and use the pre-trained optical flow algorithm to predict the optical flow from the captured object to its best matching target object, as well as the optical flow from the captured object to the entire target scene, and calculate the difference between the two to obtain the relative optical flow;

[0023] Step 11: According to the obtained relative optical flow, the active perception action is obtained through the active perception strategy network;

[0024] Step 12: Repeat steps 8-11 until there is no need to start active sensing and the number of active sensing attempts is within the specified number;

[0025] Step 13: According to the target object that best matches the grasped object obtained in step 8, if the position where the target object that best matches the grasped object is located is free, then the grasped object is placed in the target position, otherwise it is placed in the buffer zone;

[0026] Step 14: Repeat steps 1-13 until all objects are rearranged to their target states, and the task is completed within the specified number of attempts.

[0027] As a further improvement, step 2 of the present invention is specifically as follows: input the RGB image of the scene, output the bounding box set of the object, sequentially crop the original RGB image to obtain the image representation of each object bounding box, and record the object image representation in the current and target scenes as and .

[0028] As a further improvement, in step 3 of the present invention, the pre-trained visual basic model is CLIP, and step 3 specifically comprises: sequentially encoding the object image representations in the current and target scenes obtained in step 2 and , and obtain its corresponding visual features and .

[0029] As a further improvement, in step 4 of the present invention, the pre-trained grasping algorithm model is Graspnet-baseline, and step 4 is specifically as follows: input the current scene RGB-D image, and output the grasping action set in the scene Then use the grasping encoding network to encode the grasping action into grasping action features .

[0030] As a further improvement, step 5 of the present invention is specifically as follows:

[0031] Step 51: Obtain the bounding boxes of all objects in the current and target scenes according to step 2 and , calculate the 3D position coordinates of the center pixel in turn, and then encode them into the object position features in the current and target scenes through the position encoding network and ;

[0032] Step 52: The visual features of the objects in the current and target scenes obtained in step 3 , and the object position features in the current and target scenes obtained in step 51 Add together to get the visual-position features of objects in the current and target scenes, recorded as and ;

[0033] Step 53: Use the transformer cross attention mechanism to calculate the fusion features of the visual and action modalities. Specifically, the attention mechanism is expressed as:

[0034] ;

[0035] in They are query, key, value, and the Attention operator first calculates the similarity between query and key to obtain the weight, then uses softmax to normalize the weight to obtain the available weight, and finally performs weighted summation on the weight and value;

[0036] According to the attention mechanism, the visual-position features of the current scene objects are first As a query, the visual-position features of the target scene objects As key and value, the visual cross-attention features of the current and target scenes are calculated , and then grab the action features As query, the visual cross-attention features of the current and target scenes are As key and value, the cross-attention features of visual and action modalities are calculated .

[0037] As a further improvement, in step 6 of the present invention, according to the cross-attention feature obtained in step 5 , the probability values ​​logits of all grasping actions are generated through the grasping strategy network. Step 7 selects the grasping action according to the probability value logits of the grasping action obtained in step 6. During the training process, the action sampling is performed with logits as the probability; during the test process, the grasping action with the highest probability value is directly selected; then, the robot performs the corresponding grasping action.

[0038] As a further improvement, step 9 described in the present invention is specifically as follows: if the matching score is greater than the average matching score of all target objects by more than 0.12, the match is considered successful and no active perception is required.

[0039] As a further improvement, step 10 of the present invention is specifically as follows: the image representation of the object in step 2 is processed into a filling image of the same size as the scene RGB image, and masked outside the bounding box of the captured object; the filling image of the captured object is recorded as , the filled image of the target object matched by the grasped object is , use the pre-trained optical flow algorithm RAFT to predict the optical flow, arrive The object optical flow is , To the target image The global optical flow is , the difference between the two is used as the state representation of the active perception strategy, recorded as relative optical flow .

[0040] As a further improvement, in step 11 described in the present invention, the active sensing action is the continuous three-degree-of-freedom rotation angle of the end effector relative to the current posture, namely the pitch angle, roll angle, and yaw angle. By executing this action, the grasped object will be rotated to generate a new perspective for object matching.

[0041] As a further improvement, step 13 described in the present invention is specifically as follows: if the position where the matched target object is located is idle, the feature corresponding points between the grasped object and the matched target object are first extracted by the image matching algorithm SuperGlue, and then the point cloud alignment algorithm RANSAC is used for alignment to predict the target pose, and finally the grasped object is placed in the target pose; otherwise, the area in the current scene idle area that does not overlap with the target object area is calculated to form a buffer zone, and a position is randomly selected from the buffer zone to place the grasped object.

[0042] The beneficial effects of the present invention are as follows:

[0043] 1) This paper uses object-centric representation to propose a method for grasping and releasing long-sequence objects in cluttered stacking scenes based on a dual-loop strategy, which achieves efficient grasping and releasing sequence generation for long-sequence unknown objects;

[0044] 2) The present invention uses a pre-trained visual basic model to extract object visual features, match objects, learn active perception strategies, and implement a termination mechanism, showing good generalization for unknown objects;

[0045] 3) The present invention uses a pre-trained grasping algorithm model to predict candidate grasping actions in the scene, which can effectively reduce the planning action space and improve the efficiency of strategy learning;

[0046] 4) The present invention uses the transformer cross-attention mechanism to calculate the fusion features of the visual and motion features output by the pre-trained model, and evaluates the action based on the fusion features. It can achieve the alignment of visual and motion modalities with only a small amount of data and retain the generalization ability of the pre-trained model;

[0047] 5) The active perception strategy in the present invention improves the confidence of the match between the grasped object and the object in the target scene by continuously rotating the grasped object for matching, thus being more robust to perceptual noise given only the current and target scene RGB-D images, i.e., unknown object models;

[0048] 6) The present invention uses relative optical flow to characterize the uncertainty of object matching, which can provide fine-grained guidance for active perception strategies and achieve efficient active perception strategy learning;

[0049] 7) The present invention realizes the outer loop optimization of catch-and-release through reinforcement learning with task-level feedback, takes into account the optimality of the task level, avoids the misleading caused by long sequence supervision, and achieves better task success rate and efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 It is the algorithm flow chart of the present invention. DETAILED DESCRIPTION

[0051] The technical solution of the present invention is further described below by combining the accompanying drawings of the specification through specific embodiments:

[0052] The purpose of the present invention is to address the shortcomings of the grasping and releasing method of long-sequence objects in cluttered stacking scenes, such as being sensitive to perception noise and not considering the optimality of the task level. A method for grasping and releasing long-sequence objects in cluttered stacking scenes based on a double-loop strategy is proposed. The device for implementing the method of the present invention includes a camera, a robotic arm and a gripper, and a computer controller. Figure 1 It is the algorithm flow chart of the present invention, and the specific implementation method of the present invention is as follows:

[0053] A method for grasping and releasing objects in a long sequence in a cluttered stacking scene based on a dual-loop strategy, comprising:

[0054] Step 1: Given the RGB-D image of the target scene, obtain the RGB-D image of the current scene;

[0055] Step 2: Obtain all object bounding boxes in the current and target scenes through the pre-trained open set object detection algorithm and form object image representation;

[0056] Step 3: Extract visual features of all object image representations through the pre-trained visual basic model;

[0057] Step 4: Based on the current scene RGB-D image, the pre-trained grasping algorithm model is used to predict the grasping action set in the scene and encode it into grasping action features;

[0058] Step 5: Based on the obtained visual features and grasping action features of the bounding boxes of all objects in the current and target scenes, the transformer is used to implement the cross-attention mechanism of the visual and action modalities to generate cross-attention features;

[0059] Step 6: Based on the obtained cross-attention features, the probability values ​​of all grasping actions are generated through the grasping strategy network;

[0060] Step 7: Select a grasping action according to the probability value of the grasping action, and the robot executes the grasping action with the highest probability value to obtain the grasped object;

[0061] Step 8: According to the grasped object, the pre-trained visual basic model is used to match the object with all objects in the target scene, and the object in the target scene that best matches the grasped object and its matching score are obtained;

[0062] Step 9: Determine whether to start the active sensing inner loop strategy based on the obtained matching score. If active sensing is performed, execute steps 10-12; otherwise, directly execute step 13;

[0063] Step 10: According to the captured object obtained in step 7 and the target object that best matches the captured object obtained in step 8, obtain the object bounding boxes obtained in step 2 respectively, and use the pre-trained optical flow algorithm to predict the optical flow from the captured object to its best matching target object, as well as the optical flow from the captured object to the entire target scene, and calculate the difference between the two to obtain the relative optical flow, which represents the uncertainty of the current match;

[0064] Step 11: According to the obtained relative optical flow, the active perception action is obtained through the active perception strategy network;

[0065] Step 12: Repeat steps 8-11 until there is no need to start active sensing and the number of active sensing attempts is within the specified number;

[0066] Step 13: According to the target object that best matches the grasped object obtained in step 8, if the position where the target object that best matches the grasped object is located is free, then the grasped object is placed in the target position, otherwise it is placed in the buffer zone;

[0067] Step 14: Repeat steps 1-13 until all objects are rearranged to their target states, and the task is completed within the specified number of attempts.

[0068] The specific operation process of the present invention can be described as follows: a real-time RGB-D image of a desktop scene and an RGB-D image of a target scene are acquired through a depth camera Realsense L515, an open set object detection algorithm is used to detect the object bounding box, a visual basic model is used to generate visual features of all bounding boxes, and a grasping algorithm is used to generate a set of grasping actions in the scene. The cross-attention mechanism of the visual and action modes is implemented using transformers to generate cross-attention features, and then the cross-attention features are used to generate real-time grasping decisions. For the grasped object, based on the matching results of the visual basic model, it is determined whether to start the inner loop active perception strategy. If it is started, the optical flow algorithm is used to extract the relative optical flow, thereby predicting the active perception action, and rotating the grasped object to obtain a new perspective to obtain a high-confidence matching result. When the matching result meets the inner loop requirements, the object is placed according to whether the target is idle.

[0069] It should be noted that, without departing from the principles and spirit of the present invention, other method processes or operating modes obtained on the basis of the present invention without creative work are also included in the protection scope of the present invention.

[0070] In a specific implementation of the present invention, in step 1, in the RGB-D image of the initial scene, objects are stacked in a disorderly manner and there are occlusions, the target positions of the objects interfere with each other, and the posture changes of the objects in the initial and target scenes are 6 degrees of freedom.

[0071] In a specific implementation of the present invention, in step 2, the open set object detection algorithm inputs the RGB image of the scene, outputs a set of object bounding boxes, and sequentially crops the original RGB image to obtain image representations of each object bounding box, and the object image representations in the current and target scenes are recorded as and .

[0072] In step 3, the pre-trained visual base model is CLIP, which encodes the object image representations in the current and target scenes obtained in step 2 in sequence and , and obtain its corresponding visual features and .

[0073] In step 4, the pre-trained grasping algorithm model is Graspnet-baseline, and the algorithm first generates a scene point cloud from the RGB-D image of the scene through the create_point_cloud_from_rgbd_image interface of Open3D; then the Graspnet-baseline algorithm is used to generate a set of scene grasping actions , where each grasping action is represented as a 7-dimensional vector, including a 3-dimensional position vector and a 4-dimensional quaternion attitude vector; finally, the grasping action is mapped to a high-dimensional space using a grasping encoding network (three-layer MLP) to obtain the grasping action feature .

[0074] In a specific implementation of the present invention, step 5 is specifically:

[0075] Step 51: Obtain the bounding boxes of all objects in the current and target scenes according to step 2 and , using the camera external parameters and depth map, calculate the 3D position coordinates of the center pixel in turn, perform nonlinear mapping on the 3D position coordinates through trigonometric functions, and then encode them into the position features of the objects in the current and target scenes through the position encoding network and ;

[0076] Step 52: The visual features of the objects in the current and target scenes obtained in step 3 , and the object position features in the current and target scenes obtained in step 51 , Add together to get the visual-position features of objects in the current and target scenes, recorded as and ;

[0077] Step 53: Use the transformer cross attention mechanism to calculate the fusion features of the visual and action modalities. Specifically, the attention mechanism is expressed as:

[0078] ;

[0079] in They are query, key, value, and the Attention operator first calculates the similarity between query and key to obtain the weight, then uses softmax to normalize the weight to obtain the available weight, and finally performs weighted summation on the weight and value;

[0080] According to the attention mechanism, the visual-position features of the current scene objects are first As a query, the visual-position features of the target scene objects As key and value, the visual cross-attention features of the current and target scenes are calculated , and then grab the action features As query, the visual cross-attention features of the current and target scenes are As key and value, the cross-attention features of visual and action modalities are calculated .

[0081] In a specific implementation of the present invention, step 6 is based on the cross-attention feature obtained in step 5 , the probability values ​​​​logits of all grasping actions are generated through the grasping strategy network (three-layer MLP).

[0082] In a specific implementation of the present invention, step 7 selects a grasping action according to the probability value logits of the grasping action obtained in step 6, and in the training process, the action sampling is performed with logits as the probability; in the testing process, the grasping action with the highest probability value is selected; then, the robot performs the corresponding grasping action;

[0083] In a specific implementation of the present invention, step 9 determines whether the matching score is greater than the average matching score of all target objects by more than 0.12. If so, it is considered that the match is successful and no active perception is required.

[0084] In a specific implementation of the present invention, step 10 is specifically: processing the image representation of the object in step 2 into a filled image of the same size as the scene RGB image, and masking it outside the bounding box of the captured object; the filled image of the captured object is recorded as , the filled image of the target object matched by the grasped object is , use the pre-trained optical flow algorithm RAFT to predict the optical flow, input two padded images, and output an image representing the optical flow. The image size is consistent with the input padded image size, and each pixel represents the change vector between the two images; record arrive The object optical flow is , To the target image The global optical flow is ; If the match has high confidence, the optical flows of the two should be similar, so the difference between the two optical flows reflects the uncertainty of the match and can be used as a state representation of the active perception strategy, recorded as relative optical flow .

[0085] In a specific implementation of the present invention, the active sensing action described in step 11 is the continuous three-degree-of-freedom rotation angle of the end effector relative to the current posture, namely, pitch angle, roll angle, and yaw angle. By performing this action, the grasped object will be rotated to generate a new perspective for object matching.

[0086] In a specific implementation of the present invention, the active sensing strategy in step 12 is run for a maximum of 5 times, and the final matching result is retained after 5 times.

[0087] In a specific implementation of the present invention, in step 13, if the position where the matched target object is located is idle, the feature corresponding points between the grasped object and the matched target object are first extracted by the image matching algorithm SuperGlue, and then the point cloud alignment algorithm RANSAC is used to perform alignment to predict the target pose, and finally the grasped object is placed in the target pose; otherwise, the area in the idle area of ​​the current scene that does not overlap with the target object area is calculated to form a buffer zone, and a position is randomly selected from the buffer zone to place the grasped object.

[0088] In a specific implementation of the present invention, the maximum number of attempts to complete the task in step 14 is 40.

[0089] This method is trained by model-free reinforcement learning. Specifically, this method uses the discrete SAC algorithm to collect data and train in the simulator Pybullet. During the training process, the network parameters of the visual base model CLIP, the grasping model Graspnet-baseline, and the optical flow model RAFT are fixed. In each round, the current and target scenes contain 15 to 20 cluttered objects.

[0090] The above description is not a limitation of the present invention. It should be pointed out that, for ordinary technicians in this technical field, several changes, modifications, additions or substitutions can be made without departing from the essential scope of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A method for grasping and releasing objects in a long sequence in a cluttered stacking scene based on a dual-loop strategy, characterized in that: include: Step 1: Given the RGB-D image of the target scene, obtain the RGB-D image of the current scene; Step 2: Obtain all object bounding boxes in the current and target scenes through the pre-trained open set object detection algorithm and form object image representation; Step 3: Extract visual features of all object image representations through the pre-trained visual basic model; Step 4: Based on the RGB-D image of the current scene, the pre-trained grasping algorithm model is used to predict the grasping action set in the scene and encode it into grasping action features; Step 5: Based on the obtained visual features and grasping action features of all objects in the current and target scenes, the transformer is used to implement the cross-attention mechanism of the visual and action modalities to generate cross-attention features; Step 6: Based on the obtained cross-attention features, the probability values ​​of all grasping actions are generated through the grasping strategy network; Step 7: Select a grasping action according to the probability value of the grasping action, execute the grasping action with the highest probability value, and obtain the grasped object; Step 8: According to the grasped object, the pre-trained visual basic model is used to match the object with all objects in the target scene to obtain the target object that best matches the grasped object and its matching score; Step 9: Determine whether to start the active sensing inner loop strategy based on the obtained matching score. If the match is successful, no active sensing is required and step 13 is directly executed; if active sensing is performed, steps 10-12 are executed; Step 10: According to the captured object obtained in step 7 and the target object that best matches the captured object obtained in step 8, obtain the object bounding boxes obtained in step 2 respectively, and use the pre-trained optical flow algorithm to predict the optical flow from the captured object to its best matching target object, as well as the optical flow from the captured object to the entire target scene, and calculate the difference between the two to obtain the relative optical flow; The image representation of the object in step 2 is processed into a padded image of the same size as the scene RGB image, and masked outside the bounding box of the grasped object; the padded image of the grasped object is recorded as The filled image of the target object that the captured object matches is Use the pre-trained optical flow algorithm RAFT to predict optical flow. arrive The object optical flow is To the target image I g The global optical flow is If the match has high confidence, the optical flows of the two should be similar, so the difference between the two optical flows reflects the uncertainty of the match. The difference between the two is used as the state representation of the active perception inner loop strategy, recorded as relative optical flow Step 11: Based on the relative optical flow obtained, an active perception action is obtained through the active perception inner loop strategy network; the active perception action is the continuous three-degree-of-freedom rotation angle of the end effector relative to the current posture, namely, the pitch angle, the roll angle, and the yaw angle. By executing this action, the grasped object will be rotated to generate a new perspective for object matching; Step 12: Repeat steps 8-11 until there is no need to start active sensing and the number of active sensing attempts is within the specified number; Step 13: According to the target object that best matches the grasped object obtained in step 8, if the position where the target object that best matches the grasped object is located is free, then the grasped object is placed in the target position, otherwise it is placed in the buffer zone; Step 14: Repeat steps 1-13 until all objects are rearranged to their target states, and the task is completed within the specified number of attempts.

2. The method for grasping and releasing objects in a long sequence in a cluttered stacking scene based on a dual-loop strategy according to claim 1 is characterized in that: The step 2 is specifically as follows: input the RGB image of the scene, output the bounding box set of the object, cut the original RGB image in sequence to obtain the image representation of each object bounding box, and record the object image representation in the current and target scenes as and 3. The method for grasping and releasing objects in a long sequence in a cluttered stacking scene based on a dual-loop strategy according to claim 2 is characterized in that: In the step 3, the pre-trained visual basic model is CLIP, and the step 3 is specifically: sequentially encoding the object image representations in the current and target scenes obtained in step 2 and Get its corresponding visual features and 4. The method for grasping and releasing objects in a long sequence in a cluttered stacking scene based on a dual-loop strategy according to claim 1, 2 or 3, characterized in that: In step 4, the pre-trained grasping algorithm model is Graspnet-baseline, and step 4 is specifically as follows: input the current scene RGB-D image, output the grasping action set {a k } k＝1，...，L Then use the grasping encoding network to encode the grasping action into grasping action features 5. The method for grasping and releasing objects in a long sequence in a cluttered stacking scene based on a dual-loop strategy according to claim 4 is characterized in that: The step 5 is specifically as follows: Step 51: Obtain the bounding boxes of all objects in the current and target scenes according to step 2 and The 3D position coordinates of the central pixel are calculated in sequence, and then encoded into the object position features in the current and target scenes through the position encoding network. and Step 52: The visual features of objects in the current and target scenes obtained in step 3 and the object position features in the current and target scenes obtained in step 51 Add together to get the visual-position features of objects in the current and target scenes, recorded as and Step 53: Use the transformer cross attention mechanism to calculate the fusion features of the visual and action modalities. Specifically, the attention mechanism is expressed as: Attention(Q,K,V)=softmax(QK T )V; Where Q, K, and V are query, key, and value respectively. The Attention operator first calculates the similarity between query and key to obtain the weight, then normalizes the weight using softmax to obtain the available weight, and finally performs weighted summation on the weight and value. According to the attention mechanism, the visual-position features of the current scene objects are first As a query, the visual-position features of the target scene objects As key and value, the visual cross-attention features of the current and target scenes are calculated Then grab the action features As query, the visual cross-attention features of the current and target scenes are As key and value, the cross-attention features of visual and action modalities are calculated 6. The method for grasping and releasing objects in a long sequence in a cluttered stacking scene based on a dual-loop strategy according to claim 5, characterized in that: In step 6, according to the cross attention feature obtained in step 5 The probability values ​​logits of all grasping actions are generated through the grasping strategy network. The step 7 selects the grasping action according to the probability value logits of the grasping action obtained in step 6. During the training process, the action sampling is performed with logits as the probability; during the test process, the grasping action with the highest probability value is directly selected; then, the robot executes the corresponding grasping action.

7. The method for grasping and releasing objects in a long sequence in a cluttered stacking scene based on a dual-loop strategy according to claim 1 or 2 or 3 or 5 or 6, characterized in that: The step 13 is specifically as follows: if the position where the matched target object is located is idle, the feature corresponding points between the grasped object and the matched target object are first extracted by the image matching algorithm SuperGlue, and then the point cloud registration algorithm RANSAC is used for registration to predict the target pose, and finally the grasped object is placed in the target pose; otherwise, the area in the idle area of ​​the current scene that does not overlap with the target object area is calculated to form a buffer zone, and a position is randomly selected from the buffer zone to place the grasped object.

Images