Man-machine cooperation online task allocation method for complex dynamic assembly scene
Through the online task allocation method combined with GPT model and robot self-learning, the accuracy and flexibility of task allocation in complex assembly tasks are solved, dynamic task allocation is realized, and the efficiency and resource utilization of human-computer collaborative assembly are improved.
Patent Information
- Application Number
- CN202411331607.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-24
- Publication Date
- 2024-12-03
AI Technical Summary
The existing technology is difficult to achieve accurate and efficient task allocation in complex and changeable practical assembly tasks, ignoring the individual differences between robots and operators and the complexity of the on-site environment, making it difficult to meet actual needs and affecting collaboration. Assembly efficiency and quality.
Adopt a human-computer collaborative online task allocation method for complex dynamic assembly scenarios. By obtaining product assembly action sequence, task complexity evaluation, online segmented insertion learning and update task evaluation sets, combined with pre-trained model GPT, the robot's job is quantified in real time. Capability and environment perception to achieve automated and dynamic allocation of tasks.
Improve the accuracy and flexibility of task allocation, promote the harmony and efficiency of human-computer collaboration, improve the utilization rate and overall efficiency of assembly resources, and dynamically adjust task evaluation and adapt The robot's capabilities have changed significantly, which has significantly improved production and assembly efficiency.
Smart Images

Figure CN119057788A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of robot collaborative assembly and relates to a human-machine collaborative online task allocation method for complex dynamic assembly scenarios. Background Art
[0002] In the field of modern manufacturing, traditional industrial robots are not competent for the multi-variety small-batch production model and cannot independently complete all assembly operations under this production model. Collaborative robots that can integrate with humans are needed to work together with human operators on site to complete assembly tasks. Task allocation is one of the key technologies in human-machine collaborative assembly. The core challenge lies in how to accurately and efficiently allocate tasks to robots and human operators to promote a highly coordinated, safe and smooth collaborative process.
[0003] At present, task allocation strategies are mainly divided into data-driven strategies and knowledge-driven strategies. Data-driven strategies, with their powerful data analysis capabilities, such as machine learning, data mining, and statistical analysis, extract behavioral insights and trends from massive data. Knowledge-driven strategies focus on relying on expert wisdom and established rules, such as expert systems and complexity models, to provide decision support for task allocation. Among the existing knowledge-driven strategies, the allocation model based on task complexity is particularly eye-catching. It conducts a refined assessment of tasks by comprehensively considering the complexity of the task itself, the flexibility preference of operators, and the balance of task allocation (i.e., task pressure index); however, this model is unable to cope with complex and changeable actual assembly tasks. Its idealized task execution assumptions ignore the individual differences between robots and operators and the complexity of the on-site environment, resulting in task allocation results that are often difficult to meet actual needs, affecting the efficiency and quality of collaborative assembly.
[0004] In addition, traditional models mostly focus on static task allocation, that is, the allocation is completed offline before the task is executed, and the real-time changes of environmental factors during the online execution process are not fully considered. In a highly collaborative assembly environment, this static allocation method is particularly rigid, limiting the flexibility and work efficiency of operators; at the same time, it also ignores the growth of the robot's self-learning ability, resulting in task allocation failing to fully utilize the robot's potential, causing idle and wasteful resources. Summary of the invention
[0005] The purpose of the present invention is to provide an online task allocation method for human-machine collaboration for complex dynamic assembly scenarios, which solves the problem that the prior art is based on idealized task execution assumptions, focuses on static task allocation, and is difficult to cope with complex and changeable actual assembly tasks.
[0006] The technical solution adopted by the present invention is a human-machine collaborative online task allocation method for complex dynamic assembly scenarios, which is implemented according to the following steps:
[0007] Step 1: Obtain product assembly action sequence;
[0008] Step 2: Select the current task;
[0009] Step 3: Task complexity evaluation: taking the current task attributes as input and combining them with the task complexity calculation formula, evaluate whether the current task can be completed;
[0010] Step 4: Process the current task. After the evaluation is completed, if the current task can be completed, the robot will perform the assembly to assist the operator in completing the current task. After this task is completed, the next task will be selected according to step 2.
[0011] Step 5: Process the pending assembly task and determine whether the robot can perform it based on the evaluation results;
[0012] Step 6: Segmented interpolation learning: set up online segmented interpolation learning strategy, the robot creates simulation scenes online, and performs self-training of the assembly model;
[0013] Step 7: Update the task evaluation set;
[0014] Step 8: Intent change processing.
[0015] The beneficial effect of the present invention is that the method not only takes into account the dynamics and complexity of the assembly scene, but also deeply explores the impact of the robot's autonomous learning ability on task allocation. By quantifying the robot's operating ability in real time and combining dynamic environment perception and prediction technology, the automatic and dynamic allocation of assembly tasks is realized, which not only improves the accuracy and flexibility of task allocation, but also promotes the harmony and efficiency of human-machine collaboration, which is specifically manifested in the following aspects:
[0016] 1) In the core process of dynamic task allocation, the method of the present invention cleverly integrates the operator's assembly intention as a key consideration factor, and at the same time uses the pre-trained model GPT to achieve in-depth analysis and detailed division of assembly tasks.
[0017] 2) The method of the present invention further refines the task evaluation results by improving the complexity evaluation process. For uncertain tasks, an online simulation assembly test link is introduced to effectively improve the accuracy of the task complexity evaluation results and lay a solid foundation for the reasonable allocation of subsequent tasks.
[0018] 3) The method of the present invention has strong adaptive characteristics and can monitor and record the success rate of the robot's task execution in real time, and use this as a feedback mechanism to dynamically adjust and update the task evaluation set, ensure the timeliness and accuracy of the model evaluation results, flexibly adapt to changes in the robot's own capabilities, and significantly improve the utilization rate and overall efficiency of production and assembly resources.
[0019] 4) In the face of the complex problem of dynamic allocation of human-machine collaborative assembly tasks, the accuracy and efficiency of task allocation can be significantly improved by improving the traditional task complexity evaluation and introducing diversified evaluation factors to decompose and optimize the task sequence. BRIEF DESCRIPTION OF THE DRAWINGS
[0020] Figure 1 It is the overall flow chart of the method of the present invention;
[0021] Figure 2 is the human-machine collaborative assembly intention in the method of the present invention;
[0022] Figure 3 is an assembly task sequence category diagram in the method of the present invention;
[0023] Figure 4 It is a task sequence diagram generated by the task node and adjacency list method in the method of the present invention;
[0024] Figure 5 is a task attribute composition diagram in the method of the present invention;
[0025] Figure 6 is an improved task complexity evaluation flow chart in the method of the present invention;
[0026] Figure 7 It is a flow chart of online simulation scenario in the method of the present invention;
[0027] Figure 8 It is a flow chart of updating task evaluation set in the method of the present invention;
[0028] Fig. 9 is a pen assembly task decomposition diagram of method embodiment 1 of the present invention;
[0029] Fig.10 The Gantt chart of the pen assembly task allocation in the method embodiment 1 of the present invention;
[0030] Fig.11 is the continuous training reward curve of Example 3 of the method of the present invention;
[0031] Fig.12 is the segmented first training reward curve of Example 3 of the method of the present invention;
[0032] Fig.13 is the segmented second training reward curve of Example 3 of the method of the present invention;
[0033] Fig.14 It is the segmented third training reward curve of Example 3 of the method of the present invention. DETAILED DESCRIPTION
[0034] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments.
[0035] The core of the method of the present invention is to solve the problem of dynamic task allocation in the actual assembly process of human-machine collaboration. An online human-machine collaborative task allocation method integrating robot self-learning ability is proposed. On the basis of the traditional static task complexity model, an online dynamic task allocation mechanism is constructed. This mechanism optimizes the evaluation process of task complexity by deeply integrating the operator's intention and realizes the refined division of task evaluation results. Furthermore, an assembly simulation scenario is constructed online to perform real-time simulation on doubtful tasks to enhance the accuracy of decision-making.
[0036] In order to ensure the continuous optimization of the task allocation process, the method of the present invention designs a set of task evaluation sets to update the task evaluation results in real time to ensure the rationality and efficiency of the allocation; at the same time, with the help of the pre-trained model GPT, the task allocation is refined to the action sequence level, achieving more accurate task allocation. More importantly, in view of the limitations within the scope of the robot's capabilities, the method of the present invention introduces an "online gap learning" strategy, which enables the robot to automatically learn and adapt to tasks that it could not have completed independently during the assembly process, thereby achieving more perfect human-machine collaboration.
[0037] Figure 1 This is an overall flow chart of the complex task dynamic allocation of the method of the present invention. For specific steps, see steps 1 to 8. Figure 2 This is a simplified human-machine collaborative assembly flow chart. The a1, a2, and a3 in the figure represent the three states of the robot assisting in completing the current task, assisting in completing the task to be assembled, and intermittent learning, respectively, and include the operator, robot, and assembly task sequence. The assembly task consists of subtasks P1 to Pn. Parts are assembled from T1 to Tn. At a certain moment during this period, the operator performs a task, and the robot performs the operation according to the results of the dynamic task allocation model. Figure 2 At time T1, the operator intends to perform task P1, and the robot selects to assist the operator to complete the current task P1 through the task allocation model, and the operator takes a rest to adjust; at time T2, the operator intends to perform task P2, and the robot evaluates that it cannot complete P2, and executes task P3 to be assembled; at time T3, the operator intends to perform task P4, and the robot assists in completing task P4; at time T4, the robot completes the current task and the task to be assembled, performs intermittent learning, and the operator completes task P5.
[0038] Reference Figure 1 The human-machine collaborative online task allocation method of the present invention is oriented to complex dynamic assembly scenarios and is implemented according to the following steps:
[0039] Step 1: Get the product assembly action sequence.
[0040] Reference Figure 3 The assembly task sequence is divided into assembly task level, process level, work step level and action level. The assembly task level refers to the assembly task sequence between different products. The process level is the assembly task sequence determined by the assembly process file for a certain assembly product. The work step level is the task sequence obtained by the operator by dividing the process level tasks according to the operating experience. The action level is the assembly operation required to perform a certain work step in the process level.
[0041] Based on the product assembly process file information, the assembly task sequence is initially obtained. On this basis, the pre-trained model GPT is used to decompose the process and steps into the action level.
[0042] Construct task nodes and set node attributes. The node attribute information includes action name, parent process, step name, complexity evaluation attributes, execution times, evaluation results, robot execution success rate, predecessor task, successor task, etc., and use chain storage to realize digital storage of assembly task information. According to the assembly task sequence, the adjacency list method is used to generate the task sequence diagram. The node information and the adjacency list are as follows: Figure 4 shown.
[0043] The specific operation process is as follows:
[0044] First, the basic information of the robot is input into the GPT model (i.e., the pre-trained model GPT3.5), including information such as degrees of freedom, payload, arm span, end effector, and its basic capabilities.
[0045] Secondly, clarify the basic actions that the robot can perform. These operations are similar to the basic operations of operators when completing assembly tasks, usually including positioning, fastening, press-fitting, adjusting, fixing, clamping, sealing, filling, handling and grasping;
[0046] Then, based on the robot's own capabilities, design the robot's basic operations, including gripper opening, gripper closing, search action, waiting action, reset action, move and rotate action, and explain the basic principles and goals of each action;
[0047] Finally, the model input and output formats are set, and the model output constraints are determined. Taking the reducer end cover robot collaborative assembly task as an example, the prompt word construction of task decomposition is completed.
[0048] The contents of the prompt words mainly include: task objectives, basic attributes of the robot, basic capabilities of the robot, input and output formats, and examples.
[0049] A1. Mission objective: Set the virtual identity to be a seven-joint Sawyer robot processor, whose main function is to divide the actions of the tasks input by the operator;
[0050] A2. Basic properties of the robot: The robot's end effector is a two-finger gripper, with a body weight of 19kg, an arm span of 1260mm, a payload of 4kg, and a repeatability of plus or minus 0.1mm. Common assembly tools such as wrenches, screwdrivers, and torque wrenches are used in the assembly scene.
[0051] A3. Basic capabilities of the robot, pre-set actions and functions that the robot can achieve:
[0052] (1) Gripper closing action: The main function is to connect the robot to the grasped object by closing the gripper, and to achieve subsequent object movement operations without the need for task information;
[0053] (2) Mobile action: It mainly realizes the function of moving the robot to the specified position through its own motion solution. The required task information is the name of the moving target point;
[0054] (3) Gripper opening action: After the robot completes the grasping and moves the object through the movement of the robot, the object reaches the specified position and separates from the robot, and no task information is required;
[0055] (4) Waiting action: It mainly refers to the operation of the robot standing still and waiting, without any input information;
[0056] (5) Reset action: It is mainly used to reset the robot's posture and restore it to its initial state without the need for input information;
[0057] (6) Search action: mainly the robot uses visual point cloud and other technologies to realize the search function of the target object in the scene, and the task information is the name to be searched;
[0058] (7) Insertion action: After the robot grabs the parts, it inserts the grabbed parts into the parts to be assembled;
[0059] (8) Rotational action: It mainly involves the rotation of parts during the assembly process.
[0060] A4. Input and output formats,
[0061] For the robot processor, when the input starts with "Task:", the task after ":" needs to be decomposed. First, determine whether tool assistance is needed. If the corresponding tool for assembly needs to be output, decompose it in combination with the pre-set actions, obtain the information required for the action, refer to the output format of the case below, and prohibit the use of undefined actions. In addition, there is no need to modify the task requirements too much, and a unified name expression is used. There can be multiple task requirements for the search action.
[0062] A5. Example,
[0063] Input task: Assemble the reducer end cover.
[0064] Output:
[0065] Execution tasks: assembly tool wrench + search action + moving action (reducer end cover) + clamp closing action + moving action (assembly position) + insertion action + rotation action + clamp opening action.
[0066] Here are the steps:
[0067] Execute the search action to find "reducer end cover" and "assembly position". Task requirements: "reducer end cover", "assembly position".
[0068] Execute the move action and move to the position of "reducer end cover". Task requirement: "reducer end cover".
[0069] The gripper closes and grabs the "reducer end cover".
[0070] Execute the moving action to move the gripper with the "reducer end cover" to the "assembly position". Task requirement: "assembly position".
[0071] Perform the insertion action to insert the captured reducer end cover into the component to be assembled.
[0072] Perform a rotation action to rotate the reducer end cover according to assembly requirements.
[0073] Execute the jaw opening action, complete the assembly of the reducer end cover, and release the jaws.
[0074] Step 2: Select the current task.
[0075] When assigning assembly tasks, the operator's intention is introduced as a key consideration. Specifically, when faced with multiple tasks to be assigned, the operator's current intention will be identified and determined first, that is, the task that the operator currently wants to complete or focus on first; then, this intention will be used as the first choice for the current task and assigned to the robot for auxiliary execution;
[0076] In addition, when the robot completes the operation, if the operator's intention has not changed, the adjacent task in the order of completing the task will be selected as the current task; otherwise, based on the principle of priority mutual assistance in the assumption conditions, the changed operator's assembly intention will be used to update the current task.
[0077] Step 3: Task complexity evaluation
[0078] In view of the shortcomings of the traditional complexity evaluation process, this step first takes all the relevant attributes of the current task as input during the evaluation process. These relevant attributes include but are not limited to the inherent attributes of the parts, the way of task assembly and connection, and safety design. The task attribute components are as follows: Figure 5 The specific process is:
[0079] 3.1) Construct a complexity evaluation table for task assignment, as shown in Table 1.
[0080] Table 1. Complexity evaluation in task allocation
[0081]
[0082] 3.2) Calculate the task complexity by combining the complexity of the assembly parts’ own attributes, installation complexity, connection method complexity, and safety factor complexity, and calculate it using formula (1):
[0083]
[0084] Among them, L self is the component complexity score, j s L is the attribute personality of the component. con is the component installation complexity score, j c L is the attribute personality of the component. join is the complexity score of component connection mode, j j L is the attribute personality of the component. safe is the complexity score of component safety factors, j s It is the personality of the component related attributes.
[0085] 3.3) Calculate the component assembly complexity score, that is, calculate the evaluation result according to the calculation formula of each evaluation factor and the evaluation table data, and calculate it through formula (2):
[0086]
[0087] Among them, L is the component assembly complexity score, j l For the properties of the components,
[0088] Comprehensively analyze and calculate the input attributes to obtain an accurate evaluation result of the complexity of the current task.
[0089] 3.4) Based on the evaluation results, the task allocation decisions are guided as follows: (a) When the evaluation result is greater than a certain neighborhood range of the threshold and less than the threshold, the task is considered to be relatively complex and is not recommended for the robot to complete; (b) When the evaluation result is less than a certain neighborhood range of the threshold and greater than the threshold, it is determined that the task can be barely completed by the robot; for tasks with evaluation results that are not recommended or barely executed, they are uniformly marked as doubtful tasks, which require an online simulation environment to be built and simulated for verification; (c) If the evaluation result exceeds a certain neighborhood range of the threshold, it is judged that the robot has the ability to successfully complete the task; (d) If the evaluation result is less than a certain neighborhood range of the threshold, the risk of the robot operating the task is too high and cannot be completed by the robot.
[0090] 3.5) Refine the results again to complete the exact match, where a is the threshold, a is 0.7, and the neighborhood is ±0.15. The specific situation is shown in Table 2.
[0091] Table 2. Task evaluation results
[0092] Evaluation results Not executable Not recommended to complete Barely completed Executable Evaluation value 0~a-0.15 a-0.15~a a~a+0.15 a+0.15~1.0
[0093] If the evaluation result shows that the task is executable, it means that the robot has the ability to complete the task, and the task is assigned to the robot for execution; otherwise, other alternatives are considered or further optimization adjustments are made. Figure 6 As shown in the figure, this process achieves a precise match between task complexity and robot execution capability, ensuring the rationality and efficiency of task allocation.
[0094] Step 4: Process the current task.
[0095] After the task evaluation process is successfully completed and it is confirmed that the robot has the ability to perform the current task (if step 3 confirms that it can be completed autonomously), the assembly task will be handed over to the robot to assist the operator in completing the established goals; when the robot successfully completes the current task, it will automatically screen and determine the next task to be performed according to the preset task selection logic (as described in step 2) to ensure the continuity and smoothness of the assembly process.
[0096] Step 5: Processing of pending assembly tasks.
[0097] When the robot confirms after evaluation that it cannot directly assist the operator in completing the currently assigned task (if it is confirmed in step 3 that it is difficult to complete the task autonomously), the automatic scheduling mechanism for the task to be assembled is started:
[0098] First, according to the preset assembly task sequence, other assembly tasks that can be processed in parallel are selected as alternatives. Then, for these alternative tasks, the task attributes are input one by one into the optimized and improved task complexity evaluation process for comprehensive analysis and evaluation. Through this refined evaluation process, it is possible to accurately determine whether the robot is capable of performing these tasks. Based on the evaluation results, the best decision is made to assign the robot the assembly task that best suits its current capability level, ensuring the continuity and efficiency of the assembly process.
[0099] Step 6: Study in sections with breaks.
[0100] When the robot is unable to directly perform the tasks in the first two situations (i.e., the current task in step 4 and step 5 and the task to be assembled), in order to maximize the resource utilization of the assembly system and promote the continuous growth of the robot's skills, an online segmented interpolation learning strategy is introduced. The robot creates simulation scenes online and performs self-training of the assembly model.
[0101] Creating a simulation scene online means that in the process of human-machine collaboration, in order to consider the impact of the distribution of parts in the actual assembly scene on the robot's simulated assembly effect, the parts in the scene are identified and located online, and the parts are imported into the simulation platform to build a simulation scene for robot reinforcement learning assembly.
[0102] The online creation process of the simulation scene is divided into three parts: offline database construction, offline simulation scene construction, and online simulation scene construction. Figure 7 As shown, the specific process is:
[0103] 6.1) Offline library building: Before offline library building, it is necessary to build a 3D CAD model library of the parts used in the product assembly process offline in advance, so that during online simulation, after matching with the scene information, the 3D CAD model of the parts to be assembled can be imported into the Gazebo simulation scene. The specific process is: first, use SolidWorks software to model the parts and tools used in the assembly task; then, use the sw_urdf_exporter plug-in to export the 3D CAD model to the required urdf type and integrate it into the file library. The file library contains the launch file for Gazebo simulation scene call, the model geometry STL file, and the XML file used to describe the model's geometry, joints, sensors, and other related parameters.
[0104] 6.2) Offline simulation scene construction: In view of the complexity and diversity of actual assembly scenes, it is difficult to directly map from real scenes to simulation scenes. In order to reduce the complexity of online creation of simulation scenes, three-dimensional modeling of fixed objects in the assembly scene, such as toolboxes, robots, and assembly workbenches, can be pre-imported into the simulation scene before online scene construction. The specific steps are: first, create an empty Gazebo simulation scene, and use the official Sawyer_SDK to import the Sawyer robot model; then, according to the offline library building steps, build the corresponding urdf file library of the object, import the component model through the launch file interface, and adjust the model's posture; finally, save the offline simulation scene as a .world file.
[0105] 6.3) Online simulation scene construction: In the actual dynamic assembly process, the tasks that need online simulation verification can be determined through the complex dynamic task allocation process constructed in the previous article. Since the assembly objects, tools, etc. and their postures in the actual assembly process cannot be determined offline in advance, they need to be obtained through on-site machine vision recognition and positioning. The specific process is: when a certain assembly task needs to be verified online, first, the posture information of the parts in the assembly scene is obtained through the recognition and positioning module, and the information of the parts to be assembled obtained on-site is saved in the local file Pose.txt in the name-position-posture format; then, by retrieving the part names identified in the actual scene in the urdf model library constructed offline, the matching urdf model file is obtained; finally, according to the SpawnModel service type in the ServiceProxy function provided by ROS, a service communication node is generated. Through this service communication node, combined with the obtained part position and posture information in the actual scene, the three-dimensional model of the part in the actual scene is imported into the offline simulation scene pre-built in the previous article.
[0106] Through the above steps, the online creation of the simulation scene of the actual robot scene is completed. At this stage, the robot uses the gap time waiting for the assembly task to be completed to perform segmented deep reinforcement learning training. In this way, the robot can continuously improve itself and enhance its ability to cope with and execution efficiency in various complex assembly scenarios. Intermittent learning not only effectively avoids the idle waste of assembly system resources, but also provides practical opportunities for the iterative upgrade of robot skills.
[0107] Step 7: Update the task evaluation set.
[0108] like Figure 8As shown in the figure, based on the detailed analysis of the task evaluation results, the tasks are clearly divided into three categories: "unexecutable", "doubtful" and "executable", and a set of evaluation results is constructed accordingly. The evaluation result set is arranged in ascending order according to the score value, so as to intuitively show the difficulty and feasibility of each task.
[0109] For "doubtful" tasks, a dynamic verification mechanism is set up: when the actual successful execution rate of such tasks reaches 90%, it is considered that the robot's capabilities have been significantly improved to a level sufficient to cope with the task, and it is then automatically migrated from the "doubtful" set to the "executable" set. This mechanism ensures the timeliness and accuracy of task allocation, and also promotes the effective use of robot capabilities; when evaluating the tasks subsequently, refer to both the evaluation results of the task set and the evaluation results given by the complexity model, and take the maximum value of the two as the final evaluation of the task, thereby fully considering the combined influence of multiple factors and further improving the comprehensiveness and accuracy of task evaluation.
[0110] In addition, in order to continuously optimize the task allocation strategy, a mechanism is set up to select and upgrade tasks from the "unexecutable" task set: by regularly reviewing the highest-scoring tasks in the set and adding them to the "doubtful" task set for further review based on the latest evaluation results of the robot's capabilities. In this way, as the robot's capabilities continue to increase, more and more tasks will be considered executable, thereby gradually reducing the number of "unexecutable" tasks and achieving continuous optimization and efficient advancement.
[0111] Step 8: Intent change processing,
[0112] During this process, the robot continuously and periodically receives inputs of the operator's intentions during assembly, so as to keenly capture changes in the operator's intentions. Once a change in intention is detected, it is regarded as a signal that the operator is about to perform a new task, and the updated operator's assembly intentions and the latest assembly sequence are automatically re-evaluated as key inputs. At the same time, every time the robot successfully completes the current task or is about to proceed to the next assembly task, it will check again whether the operator's intentions have been adjusted, and flexibly adjust accordingly to reselect the current priority task. This series of dynamic response mechanisms ensures the high flexibility and accuracy of the human-machine collaborative assembly process.
[0113] The innovation of the present invention is: first, the operator's assembly intention is introduced as a key factor in task allocation, so that the robot can respond to the operator's needs in real time and give priority to assisting it to complete the task of current concern, thereby enhancing the flexibility and tacit understanding of human-machine collaboration. Secondly, the advanced pre-training model GPT is used to achieve deep decomposition of assembly tasks, and complex processes and steps are refined to the action level, providing more accurate data support for subsequent complexity evaluation and dynamic adjustment. On this basis, the traditional task complexity evaluation process is improved, and the accuracy and reliability of the evaluation results are improved by refining evaluation indicators and introducing online simulation assembly tests. In order to maximize the resource utilization of the assembly system and promote the continuous improvement of robot skills, an online segmented interpolation learning strategy is introduced. The robot creates a simulation scene online and conducts autonomous reinforcement learning training. At the same time, a set of dynamically updated task evaluation set mechanisms is designed to adjust the priority and feasibility of task allocation in real time according to the success rate and ability changes of the robot's task execution.
[0114] Example 1
[0115] Evaluation of the task complexity of the secondary reducer. Taking the assembly of the secondary reducer as an example, the parameters of the components are given, as shown in Table 3.
[0116] Table 3. Detailed list of components for secondary reducer assembly
[0117]
[0118]
[0119] Based on the above method process of the present invention, it is implemented according to the following steps:
[0120] Step 1: Get the product assembly action sequence.
[0121] The task sequence diagram is generated using the adjacency linked list method according to the assembly task sequence.
[0122] Step 2. Select the current task: From the perspective of giving full play to the collaborative characteristics of human-machine collaboration, the operator's intention is introduced in the process of allocating assembly tasks; and the operator's intention is taken as the current task. The robot gives priority to assisting the operator to complete the current assembly task.
[0123] Step 3: Task complexity evaluation: Take the current task attributes as input and combine them with the task complexity calculation formula to evaluate whether the current task can be completed.
[0124] Step 4: Process the current task.
[0125] After the evaluation is completed, if the current task can be completed, the robot will perform the assembly to assist the operator in completing the current task; after this task is completed, select the next task according to step 2.
[0126] Step 5: Processing of pending assembly tasks: When the robot evaluates that it cannot assist the operator in completing the current task, it selects other parallel pending assembly tasks based on the assembly task sequence; inputs its task attributes into the improved task complexity evaluation process, and determines whether the robot can execute it based on the evaluation results.
[0127] Step 6: Interval learning: When the robot cannot perform the first two situations, in order to avoid wasting resources in production and assembly, the robot performs intermittent learning; while waiting for the assembly task to be completed, the robot improves its own capabilities through reinforcement learning training.
[0128] Step 7: Update the task evaluation set.
[0129] According to the detailed evaluation results, the evaluation results of the tasks are divided into unexecutable, doubtful and executable, and an evaluation result set is constructed and arranged in ascending order according to the score value; when the actual successful execution rate of the doubtful task reaches 90%, it is considered that the robot's ability has been improved enough to complete the task, and the task is moved to the executable task set; in the subsequent evaluation of tasks, the maximum value of the task set evaluation result and the complexity model evaluation result is taken as the final evaluation result of the task; at the same time, the task with the highest score in the set is selected from the unexecutable task set and included in the doubtful task set;
[0130] As the capabilities of robots increase, more tasks can be performed and fewer tasks cannot be performed.
[0131] Step 8: Intent change processing,
[0132] The operator's assembly intention is periodically obtained to detect whether the operator's intention has changed. When a change in intention is detected, it means that the operator's execution task has changed, and the operator's assembly intention and assembly sequence will be re-input; in addition, when the robot completes the current task or the task to be assembled, it is necessary to determine whether the operator's intention has changed and reselect the current task.
[0133] By evaluating the moving and inserting actions in each task in the reducer assembly, the assembly task evaluation results are obtained, as shown in Table 4.
[0134] Table 4. Evaluation results of reducer assembly tasks
[0135]
[0136]
[0137]
[0138] According to the evaluation results obtained in Table 4, it can be found that in the assembly process of the secondary reducer, most tasks are difficult for the robot, and further assembly simulation of the robot is required to evaluate the feasibility of the robot to perform the task.
[0139] Example 2
[0140] Since the reducer assembly task is too complicated, and the assembly situation of a certain step of the reducer is too simple and lacks sufficient representativeness, this embodiment 2 takes the human-machine collaborative assembly of a fountain pen as an example to verify the above process of the method of the present invention. The assembly task of the fountain pen includes three assembly steps: pen tip installation, pen cap installation, and button installation, which are represented as task 1, task 2, and task 3 respectively. The pen tip installation and button installation are taken as parallel tasks, and the three tasks are decomposed through GPT, such as Fig. 9 The assembly diagram input is at the action level, and then the dynamic allocation result is given, as shown in Table 5.
[0141] Table 5. Dynamic allocation results of pen assembly in Example 2
[0142]
[0143]
[0144] According to Table 5, the Gantt chart of the collaborative assembly process between operators and robots for pen assembly is given, as shown in Fig.10 As shown, it can be found that the robot and the operator can efficiently collaborate to complete the pen assembly task.
[0145] Example 3
[0146] In order to make full use of robot resources, when the robot cannot complete the assembly task, the robot sets up a simulation failure scenario online, the robot enters the intermittent learning state and trains the deep reinforcement learning algorithm model to achieve the growth of the robot's own capabilities. During this period, when the operator's intention changes, the robot ends the learning state. In order to verify the feasibility of improving the robot's capabilities through intermittent learning, a complete vertical assembly model training process is divided into three times for learning on average. During the initial training, the neural network parameters are randomly initialized, and the neural network parameters after the previous training are loaded at the beginning of each training. The cumulative reward curves for continuous training and segmented training are as follows Fig.11 , Fig.12 , Fig.13 , Fig.14 As shown, the continuous training design is 3600 rounds, and the segmented training is one training every 1200 rounds.
[0147] Depend on Fig.11 , Fig.12 , Fig.13 , Fig.14 It can be seen that the segmented training model will continue training based on the results of the previous training to a certain extent at the beginning of the next training. By comparing the average reward value of the final model, it can be found that after the same training round, the average reward of the third model of segmented learning can basically reach the result of the continuous training model. In addition, in the three segmented trainings, the average reward value at the beginning of the model increased from 50 to about 250. Therefore, the robot's online intermittent learning simulation can help improve the robot's own capabilities to a certain extent. Although this segmented learning method needs to go through an exploration phase at the beginning of each training cycle, it can still achieve basically the same performance as continuous training in the end. In this way, the robot can gradually learn and improve its performance in actual assembly tasks.
[0148] Example 4
[0149] In order to test the task decomposition of the fine-tuned GPT model, the bolt assembly task was decomposed and tested, and the output results are shown in Table 6.
[0150] Table 6. Assembly bolt examples
[0151]
[0152] It can be found that the trained GPT model can decompose the assembly task and list the task requirements, which can basically meet the subsequent task evaluation requirements.
Claims
1. A human-machine collaborative online task allocation method for complex dynamic assembly scenarios, characterized in that: Follow these steps to implement: Step 1: Obtain product assembly action sequence; Step 2: Select the current task; Step 3: Task complexity evaluation: taking the current task attributes as input and combining them with the task complexity calculation formula, evaluate whether the current task can be completed; Step 4: Processing the current task. After the evaluation is completed, it is judged that if the current task can be completed, the robot will perform the assembly to assist the operator in completing the current task. After completing this task, select the next task according to step 2; Step 5: Process the pending assembly task and determine whether the robot can perform it based on the evaluation results; Step 6: Segmented interpolation learning: set up online segmented interpolation learning strategy, the robot creates simulation scenes online, and performs self-training of the assembly model; Step 7: Update the task evaluation set; Step 8: Intent change processing.
2. The method for online task allocation for complex dynamic assembly scenarios based on human-machine collaboration according to claim 1 is characterized in that: In step 1, the specific process is: First, input the basic information of the robot into the GPT model; Secondly, identify the basic actions that the robot can perform, which are similar to the basic actions of operators when completing assembly tasks; Then, based on the robot's own capabilities, design the robot's basic operations, including gripper opening, gripper closing, search action, waiting action, reset action, move and rotate action, and explain the basic principles and goals of each action; Finally, set the format of model input and output, determine the model output constraints, and complete the construction of prompt words for task decomposition.
3. The method for online task allocation for complex dynamic assembly scenarios based on human-machine collaboration according to claim 1 is characterized in that: In step 2, the specific process is: When faced with multiple tasks to be assigned, the operator's current intention is first identified and determined, that is, the task that the operator currently wants to complete or focus on first; then, this intention is used as the first choice for the current task and is assigned to the robot for auxiliary execution; In addition, when the robot completes the operation, if the operator's intention has not changed, the adjacent task in the order of completing the task will be selected as the current task; otherwise, based on the principle of priority mutual assistance in the assumption conditions, the changed operator's assembly intention will be used to update the current task.
4. The method for online task allocation for complex dynamic assembly scenarios based on human-machine collaboration according to claim 1 is characterized in that: In step 3, the specific process is: 3.1) Construct a complexity evaluation table for task assignment; 3.2) Calculate the complexity of the task; 3.3) Calculate the component assembly complexity score; 3.4) Guide task allocation decisions based on the evaluation results; 3.5) Refine the results again to complete the accurate matching.
5. The method for online task allocation for complex dynamic assembly scenarios based on human-machine collaboration according to claim 1 is characterized in that: In step 4, the specific process is: After the task evaluation process is successfully completed and it is confirmed that the robot has the ability to perform the current task, the assembly task will be handed over to the robot to assist the operator in completing the established goals; when the robot successfully completes the current task, it will automatically screen and determine the next task to be performed based on the preset task selection logic and refer to step 2 to ensure the continuity and smoothness of the assembly process.
6. The method for online task allocation for human-machine collaboration in complex dynamic assembly scenarios according to claim 1 is characterized in that: In step 5, the specific process is: First, according to the preset assembly task sequence, other assembly tasks that can be processed in parallel are selected as candidates. Then, for these candidate tasks, the task attributes are input one by one into the optimized task complexity evaluation process to conduct a comprehensive analysis and evaluation to accurately determine whether the robot is capable of performing these tasks. Based on the evaluation results, the optimal decision is made to assign the robot the assembly task that best suits its current ability level.
7. The method for online task allocation for complex dynamic assembly scenarios based on human-machine collaboration according to claim 1 is characterized in that: In step 6, the specific process is: Creating simulation scenes online means that in the process of human-machine collaboration, the impact of the distribution of parts in the actual assembly scene on the robot simulation assembly effect is considered, the parts in the scene are identified and located online, and the parts are imported into the simulation platform to build a simulation scene for robot reinforcement learning assembly; The online creation process of the simulation scene is divided into offline database construction, offline simulation scene construction, and online simulation scene construction, specifically: 6.1) Offline library construction: First, use SolidWorks software to build 3D CAD models of parts and tools used in assembly tasks; then, use the sw_urdf_exporter plug-in to export the 3D CAD model to the required urdf type and integrate it into the file library; 6.2) Offline simulation scene construction: First, create an empty Gazebo simulation scene and use Sawyer_SDK to import the Sawyer robot model; then, according to the offline library building steps, build the urdf file library corresponding to the object, import the component model through the launch file interface, and adjust the model's posture; finally, save the offline simulation scene as a .world file; 6.3) Online simulation scene construction: When an assembly task needs to be verified online, first, the position and posture information of the parts in the assembly scene is obtained through the recognition and positioning module, and the information of the parts to be assembled obtained on site is saved in the local file Pose.txt in the format of name-position-pose; then, the names of the parts identified in the actual scene are retrieved from the urdf model library built offline, and the matching urdf model file is obtained; finally, according to the SpawnModel service type in the ServiceProxy function, a service communication node is generated, and through the service communication node, combined with the obtained part position and pose information in the actual scene, the three-dimensional model of the parts in the actual scene is imported into the offline simulation scene pre-built in the previous article.
8. The method for online task allocation for complex dynamic assembly scenarios based on human-machine collaboration according to claim 1 is characterized in that: In step 7, the specific process is: Based on the detailed analysis of the task evaluation results, the tasks are clearly divided into three categories: "unexecutable", "doubtful", and "executable". Based on this, a set of evaluation results is constructed. The evaluation result set is arranged in ascending order according to the score value, so as to intuitively display the difficulty and feasibility of each task.
9. The method for online task allocation for complex dynamic assembly scenarios based on human-machine collaboration according to claim 1 is characterized in that: In step 8, the specific process is: The robot continuously and periodically receives input of the operator's intention during assembly and captures changes in the operator's intention. Once a change in intention is detected, it is regarded as a signal that the operator is about to perform a new task, and the updated operator's assembly intention and the latest assembly sequence are then re-evaluated as key inputs. At the same time, every time the robot successfully completes the current task or is about to proceed to the next assembly task, it will check again whether the operator's intention has been adjusted, and adjust accordingly, and reselect the current priority task.
Citation Information
Cited By
Task matching method and system
CN119272070A
Data acquisition method, device and equipment for intelligent robot with body and medium
CN120422234A
Multi-scene switching scheduling method and system for intelligent robot with body
CN120972667A
Self-adaptive assembly guiding method and system based on multi-source data fusion
CN121189776A
