Sequence planning method for space truss assembly

Abstract

The embodiment of the invention provides a sequence planning method for space truss assembly, and the method comprises the steps: generating a supporting point replacement operation cost function for a space truss assembly task according to the energy consumption cost and time cost of the two-arm coordination supporting point replacement operation of the space truss assembly; generating a space truss assembly task representation model; according to the supporting point replacement operation cost function, the assembly consumption time, the mechanical arm disturbance torque and the operability of the mechanical arm, a comprehensive cost evaluation index is determined; according to the comprehensive cost evaluation index and the space truss assembly task representation model, a deep network is generated, the deep network is utilized to train a strategy network of the space truss assembly sequence, the trained strategy network of the space truss assembly sequence is obtained, and the strategy network is used for obtaining the space truss assembly sequence. According to the technical scheme provided by the embodiment of the invention, the strategy network of the space truss assembly sequence can be obtained so as to realize sequence planning of space truss assembly.

Description

Technical Field

[0001] The present invention belongs to the technical field of space truss assembly, and particularly relates to a sequence planning method for space truss assembly.

Background Art

[0002] Since the end of 2022, China's space station has officially entered the application and development stage. With the continuous deepening of China's research in the space field, the functional requirements of space facilities have been increasing, and more and more large-scale space truss structures need to be assembled, such as spliced large-aperture space telescopes, large-scale space solar power stations, etc.

[0003] Currently, because manual assembly by astronauts has high risks and great limitations, unmanned space truss assembly by space manipulators is a current research hotspot. The space truss assembly task has a long process and cumbersome assembly steps. It is difficult for a single-arm space robot to stably complete the space truss assembly task, while a two-arm space robot has strong flexibility and load capacity. Through coordinated operation of the two arms, the space truss assembly task can be completed more safely and stably. Assembly sequence planning is a key technology for intelligent unmanned assembly of space trusses. Its significance lies in detecting feasible assembly sequences and finding the optimal assembly sequence considering various constraints in the assembly process, which is beneficial to improving assembly efficiency and reducing assembly costs. For the sequence planning of space truss assembly, the existing technologies are difficult to simultaneously optimize multiple aspects such as assembly time and assembly stability, resulting in poor comprehensive performance of the assembly sequence planning results. Therefore, in order to achieve unmanned autonomous space truss assembly and optimize the sequence planning of space truss assembly, the sequence planning of space truss assembly with coordinated operation of two arms is an urgent problem to be solved.

Summary of the Invention

[0004] In view of this, the present invention provides a sequence planning method for space truss assembly, which comprehensively measures the assembly time, energy consumption, assembly stability and the operability of the manipulator to improve the comprehensive performance of the assembly sequence planning results.

[0005] An embodiment of the present invention provides a sequence planning method for space truss assembly, including:

[0006] For the space truss assembly task, generate a support point replacement operation cost function based on the energy consumption cost and time cost of the coordinated support point replacement operation of the space truss assembly;

[0007] Generate a space truss assembly task characterization model;

[0008] Determine a comprehensive cost evaluation index based on the support point replacement operation cost function, assembly consumption time, manipulator disturbance torque and the operability of the manipulator;

[0009] According to the comprehensive cost evaluation index and the spatial truss assembly task representation model, a deep network is generated. Using the deep network, the policy network of the spatial truss assembly sequence is trained to obtain the trained policy network of the spatial truss assembly sequence. The policy network is used to obtain the spatial truss assembly sequence, and the spatial truss assembly sequence includes a sequence representing the assembly order of multiple truss components, a sequence representing the responsibilities of the robotic arm, and a sequence representing the positions of the support points.

[0010] In the above method, for the spatial truss assembly task, according to the energy consumption cost and time cost of the double-arm coordinated support point replacement operation of the spatial truss assembly, a support point replacement operation cost function is generated, including:

[0011] The generated support point replacement operation cost function is:

[0012] Z = P grab + P loose + M1 + M2

[0013] Where Z represents the total cost of the support point replacement operation, P grab represents the time cost and energy consumption cost required for the support arm to grasp the assembled truss structure in the next state, P loose represents the time cost and energy consumption cost required for the support arm to release the assembled truss structure in the current state, and P grab 、P loose are affected by the truss shape and the end effector of the robotic arm; M1 represents the time cost and energy consumption cost required for the support arm to move during the support point replacement operation, and M2 represents the time cost and energy consumption cost required for the manipulator to move during the support point replacement operation.

[0014] In the above method, the method for generating the spatial truss assembly task representation model includes:

[0015] Generate the self-attributes of the truss components;

[0016] Generate the spatial truss structure connection matrix C;

[0017] Generate the assembly state representation;

[0018] Generate the assembly constraint representation, including the truss component connection constraint and the truss component assembly order constraint;

[0019] According to the self-attributes of the truss components, the spatial truss structure connection matrix C, the assembly state representation, and the assembly constraint representation, determine the spatial truss assembly task representation model.

[0020] In the above method, the self-attributes of the truss components include:

[0021] Q = [q1, q2, …, q r ​

[0022] A i =[a i ,m i

[0023] Wherein, Q represents the set of truss components, and A i represents the self-attribute of the i-th truss component, q i represents the i-th truss component. Suppose there are r truss components in total, including m connectors and h truss bars. When i ≤ m, q i represents a connector. When i > m, q i represents a truss bar, a i represents the number of interfaces of the truss component q i , and a i ≥2, m i represents the mass of the truss component q i .

[0024] In the above method, the spatial truss structure connection matrix C includes:

[0025]

[0026] Wherein, c i j = c(i, j), and C represents the connection relationship between truss components.

[0027] In the above method, generating the assembly state characterization includes:

[0028]

[0029] S T =[p1, p2, …, p r

[0030]

[0031] Wherein, W i represents the description of the assembly states of all adjacent truss components of the truss component q i . When a truss component has multiple adjacent truss components, they are sorted according to the original ID numbers of the adjacent truss components. w i1 ~w ia represent the assembly states of the a adjacent truss components of the truss component q i . 0 represents that the adjacent truss component does not exist, 1 represents that the adjacent truss component is not assembled, and 2 represents that the adjacent truss component is assembled. S T represents the assembly state of the spatial truss structure.

[0032] In the above method, the truss component connection constraints include: ​​

[0033] K i = (1 - p i )h i

[0034] where h i represents the number of adjacent truss members q i that have been assembled, and p i is consistent with the meaning of p defined in the assembly state characterization. K i ≠ 0 indicates that the truss member q i satisfies this constraint condition; i

[0035] The generation of the truss member assembly sequence constraint is:

[0036] H i = (w i1 - 1)(w i2 - 1), (i > m)

[0037] where w i1 and w i2 represent the assembly states of two adjacent truss members of the truss rod q i . The space truss member assembly sequence constraint condition H i = 0 indicates that at least one of the adjacent truss members of the truss rod q i has not been assembled.

[0038] In the above method, based on the support point replacement operation cost function, assembly consumption time, manipulator disturbance torque, and the operability of the manipulator, a comprehensive cost evaluation index is determined, including:

[0039] The determined comprehensive cost evaluation index L is:

[0040] L = γ1ε1t all + γ2ε2τ + γ3ε3ω + Z

[0041] where γ1 represents the weight of the time cost index, γ2 represents the weight of the assembly stability index, and γ3 represents the weight of the operability index; ε1 represents the normalization parameter of the time cost, ε2 represents the normalization parameter of the assembly stability, and ε3 represents the normalization parameter of the operability; t all = t d + t o , t all represents the assembly consumption time, t d represents the running time of the path planning algorithm, and t o is the time consumed by the manipulator to perform the assembly task; τ = ||r × F||, τ represents the manipulator disturbance torque, r represents the position vector from the rotation center to the force application point, and F represents the applied assembly force; ω represents the operability of the manipulator, ​$J(q)$ is the Jacobian matrix of the robotic arm when the joint angle is $q$; $Z$ represents the cost function for the support point replacement operation.

[0042] In the above method, according to the comprehensive cost evaluation index and the spatial truss assembly task characterization model, a deep network is generated, including:

[0043] Construct the state space $S$ at time $t$ t as:

[0044] $S$ t $=[s$ dis , $s$ sup , $s$ truss $

[0045] where $s$ dis is a scalar representing the duty assignment of the robotic arm. For a dual-arm task scenario, there are only two duty assignment methods. When the value is 1, it means that robotic arm 1 is the support arm and robotic arm 2 is the operating arm. When the value is 0, it means that robotic arm 2 is the support arm and robotic arm 1 is the operating arm. $s$ sup is a scalar representing the support point position, corresponding to the ID number of the truss rod. When $s$ sup $=i$, it means that the support point is at the centroid position of the truss rod with ID number $i$. $s$ truss is a scalar obtained by dimensionality reduction using the assembly state $S$ T of the spatial truss structure, representing the current state of the truss structure;

[0046] Construct the action space $A$ at time $t$ t as:

[0047] $A$ t $=[a1, a2, a3]$

[0048] where $a1$ is a scalar representing the ID number of the truss component to be assembled, $a2$ is a scalar representing whether to perform the support arm replacement operation, and $a3$ is a scalar representing the support point position selection. When $a2 = 0$ and $a3 = s$ sup , no support arm or support point replacement operation is performed. When $a2 = 1$, only the support arm replacement operation is performed without changing the support point position. When $a2 = 0$ and $a3 \neq s$ sup , only the support point position is changed without changing the support arm. When $a2 = 1$ and $a3 \neq s$ sup , both the support arm replacement operation and the support point position change are performed;

[0049] Construct the reward function $R$ as:

[0050] $R=-L$

[0051] where $L$ is the comprehensive cost evaluation index;

[0052] Construct a deep network, which includes a policy network π θ and a value network V φ , the policy network π θ takes the state space S t as the input and the action space A t as the output; the value network V φ takes the state space S t as the input and the value as the output.

[0053] In the above method, the trained policy network for the spatial truss assembly sequence is obtained by using the deep network, and the steps include:

[0054] Step 1: Obtain the connection matrix of the spatial truss structure, the initial poses of each truss component, the target poses of each truss component, the base poses of the two robotic arms, and the initial end poses;

[0055] Step 2: Let the current time step t = 1;

[0056] Step 3: Take the state space S t =[s dis , s sup , s truss at time step t as the input of the policy network π θ to obtain the action space A t =[a1, a2, a3] at time t;

[0057] Step 4: According to the assembly constraint representation, judge whether A t meets the robotic arm operation space constraint. If the judgment result is no, execute Step 3. If the judgment result is yes, then enter Step 5;

[0058] Step 5: Calculate the reward function value r t according to the reward function R;

[0059] Step 6: Combine S t , A t , r t , and S t+1 to construct an experience pool;

[0060] Step 7: Let t ← t + 1, and loop to execute Steps 3 to 7 until all truss components are traversed;

[0061] Step 8: Use the constructed experience pool and the PPO algorithm to update the parameters of the policy network π θ and the value network V φ ;

[0062] Step 9: When all the data in the experience pool has been used for training, end the training and obtain the policy network π of the spatial truss assembly sequence. θ .

[0063] After the training is completed, the policy network can, according to the input task information, output an optimal action sequence that takes into account the time cost, assembly stability, and manipulator operability, and includes the information of dual-arm coordinated operation.

Description of the Drawings

[0064] To more clearly illustrate the technical solutions of the embodiments of the present invention, the following will briefly introduce the drawings required in the embodiments. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative and laborious efforts, other drawings can also be obtained based on these drawings.

[0065] Figure 1 is the operation flowchart of a sequence planning method for spatial truss assembly provided by an embodiment of the present invention;

[0066] Figure 2 is a schematic diagram of a rectangular truss structure in the assembled state in an embodiment of the present invention;

[0067] Figure 3 is a schematic diagram of a rectangular truss structure in the assembling state in an embodiment of the present invention;

[0068] Figure 4 is a schematic diagram of a double-cube truss structure in an embodiment of the present invention;

[0069] Figure 5 is a comparison chart of the comprehensive cost of the double-cube planning task with the ant colony algorithm in an embodiment of the present invention.

Detailed Embodiments

[0070] To better understand the technical solutions of the present invention, the following will describe the embodiments of the present invention in detail with reference to the drawings.

[0071] It should be clear that the described embodiments are only some embodiments of the present invention, rather than all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present invention.

[0072] An embodiment of the present invention provides a sequence planning method for spatial truss assembly. As Figure 1 shown, the method includes the following steps:

[0073] Step 101: For the space truss assembly task, generate a support point replacement operation cost function based on the energy consumption cost and time cost of the dual-arm coordinated support point replacement operation for space truss assembly.

[0074] Specifically, considering that the support point replacement operation has the characteristics of energy consumption and time cost, the support point replacement operation cost function is generated as follows:

[0075] Z = P grab + P loose + M1 + M2

[0076] Among them, Z represents the total cost of the support point replacement operation, P grab represents the time cost and energy consumption cost required for the support arm to grasp the assembled truss structure in the next state, P loose represents the time cost and energy consumption cost required for the support arm to release the assembled truss structure in the current state, and P grab 、P loose 's values are affected by the truss shape and the fixture at the end of the manipulator; P loose represents the time cost and energy consumption cost required for the support arm to move during the support point replacement operation, and M2 represents the time cost and energy consumption cost required for the manipulator arm to move during the support point replacement operation.

[0077] Specifically, the time cost refers to the time t c consumed by the manipulator arm to complete the corresponding action, and the energy consumption cost refers to the energy E c consumed by the manipulator arm to complete the corresponding action. E c = P MA t c ,, where P MA refers to the power of the manipulator arm, P grab = ε grab1 t c + ε grab2 E c ,ε grab1 and ε grab2 are normalization parameters in cost calculation, and are valued according to manual experience. The calculation methods of P loose 、P loose and M2 are the same as the calculation method of P grab .

[0078] So far, the generation of the support point replacement operation cost function has been completed.

[0079] Step 102: Generate a space truss assembly task characterization model.

[0080] The space truss assembly task characterization model is determined by the generated truss component's own attributes, the space truss structure connection matrix C, the assembly state characterization, and the assembly constraint characterization.

[0081] In a feasible implementation, the self-attributes of the truss components are generated as follows:

[0082] Q = [q1, q2, …, q r

[0083] A i = [a i , m i

[0084] where Q represents the set of truss components, and A i represents the self-attributes of the i-th truss component, including the number of interfaces and the mass of the truss component. q i represents the i-th truss component. Suppose there are r truss components in total, including m connectors and h truss bars. When i ≤ m, q i represents a connector. When i > m, q i represents a truss bar. a i represents the number of interfaces of the truss component q i , and a i ≥ 2. m i represents the mass of the truss component q i ;

[0085] Specifically, the rectangular truss structure in the assembled state in the embodiments of the present invention is as Figure 2 shown in the figure. In the figure, the numbers represent the ID numbers of the truss components, the dots represent the connectors, and the straight lines represent the truss bars. The space truss is a square structure with a side length of 0.2 m and a mass of 1 kg, including 4 connectors and 4 truss bars. And each connector has 2 interfaces, and each truss bar has 2 interfaces. Therefore, m = 4, h = 4, r = 8, and A5 = [0.2, 1].

[0086] In a feasible implementation, a connection matrix C of the space truss structure is generated:

[0087]

[0088] The connection relationship between the truss components is represented by the connection matrix C of the space truss structure;

[0089] For example, according to Figure 2 , the connection matrix of the space truss structure in the example can be defined as:

[0090]

[0091] In a feasible implementation, an assembly state representation is generated:

[0092] To confirm the component q of the space truss unit i ​​Whether it has the assembly conditions requires understanding the assembly status of its adjacent components. Define the components q of the space truss element i The assembly status of the adjacent components is as follows:

[0093]

[0094] Among them, W i represents the description of the assembly status of all adjacent truss components of the truss component q i . When a truss component has multiple adjacent truss components, they are sorted according to the original ID numbers of the adjacent truss components. w i1 ~w ia represent the assembly status of a adjacent truss components of the truss component q i . 0 means that the adjacent truss component does not exist, 1 means that the adjacent truss component is not assembled, and 2 means that the adjacent truss component is assembled. S T represents the assembly status of the space truss structure;

[0095] To describe the assembly status, define S T to represent the assembly status of the space truss structure. S T is expressed as:

[0096] S T =[p1,p2,…,p r

[0097] Among them, p i represents whether the component q i is assembled. When q i is assembled, then p i =1; when q i is not assembled, p i =0. p i can be expressed as:

[0098]

[0099] Specifically, the rectangular truss structure in the assembled state is as shown in Figure 3 . In the figure, the numbers represent the ID numbers of the truss components, the dots represent the connectors, the straight lines represent the truss bars. The component q2 contains two adjacent truss components, namely q5 and q6. q5 has been assembled and q6 has not been assembled. Then the assembly status W i of the adjacent components of q2 is [1,0]; for example, the assembly status can be expressed as S T =[1,1,0,1,1,0,0,1].

[0100] In a feasible implementation, generate the connection constraints of the truss components:

[0101] ​Component connection constraint means that if we want to assemble truss components, it is required that truss component q i itself is in an unassembled state, and at least one of its adjacent components has been assembled. Define K i to describe this constraint, K i is defined as:

[0102] K i =(1 - p i )h i

[0103] where h i represents the number of adjacent truss components of truss component q i that have been assembled. p i has the same meaning as the p i defined by the assembly state characterization. K i ≠0 means that truss component q i satisfies this constraint condition;

[0104] In a feasible implementation, the truss component assembly sequence constraint:

[0105] The truss component assembly sequence constraint refers to the constraint on the installation sequence of different components. According to the characteristics of truss component assembly: when assembling a truss rod, at least one of the adjacent components of the truss rod is unassembled. Therefore, define H i to describe this constraint, H i is expressed as:

[0106] H i =(w i1 - 1)(w i2 - 1),(i > m)

[0107] where w i1 and w i2 represent the assembly states of two adjacent truss components of truss rod q i . The spatial truss component assembly sequence constraint condition H i =0 means that at least one of the adjacent truss components of truss rod q i is unassembled.

[0108] Step 103: Determine the comprehensive cost evaluation index based on the support point replacement operation cost function, assembly consumption time, manipulator disturbance torque, and manipulator operability.

[0109] Assembly sequence evaluation aims to select the optimal assembly sequence that meets the assembly conditions from the feasible assembly sequences according to the assembly sequence evaluation indicators. It is necessary to comprehensively consider the costs in multiple aspects, so the comprehensive cost evaluation indicators need to be determined. However, the dimensions of each aspect indicator are different, and direct comparison will lead to deviations in the results. Therefore, it is necessary to normalize these indicators so that they can be compared within the same numerical range. According to the characteristics of the truss component assembly operation, the effect of the assembly sequence is mainly affected by the assembly time, assembly stability, and operability. Considering the cost of the support point replacement operation in step 101, the comprehensive cost evaluation indicator is defined as:

[0110] L = γ1ε1t all +γ2ε2τ + γ3ε3ω + Z

[0111] where γ1 represents the weight of the time cost indicator, γ2 represents the weight of the assembly stability indicator, γ3 represents the weight of the operability indicator. These three weights need to be given based on practical situations and manual experience; ε1 represents the normalization parameter of the time cost, ε2 represents the normalization parameter of the assembly stability, ε3 represents the normalization parameter of the operability, which are used to achieve the normalization operation; t all = t d + t o t all represents the time cost indicator, t d represents the strategy deployment time, that is, the running time of the path planning algorithm, and t o is the time consumed by the robotic arm to perform the assembly task; τ = ||r × F||, τ represents the disturbing torque indicator, r represents the position vector from the rotation center to the force application point, and F represents the applied assembly force; ω represents the operability indicator, J(q) is the Jacobian matrix of the robotic arm when the joint angle is q; Z represents the support point replacement cost function.

[0112] Specifically, in the example, γ1 = 0.2, γ�2 = 0.5, and γ3 = 0.3.

[0113] So far, the generation of the comprehensive cost evaluation indicator has been completed.

[0114] Step 104: Generate a deep network based on the comprehensive cost evaluation indicator and the spatial truss assembly task characterization model. Use the deep network to train the policy network of the spatial truss assembly sequence, and obtain the trained policy network of the spatial truss assembly sequence. The policy network is used to obtain the spatial truss assembly sequence, and the spatial truss assembly sequence includes a sequence representing the assembly order of multiple truss components, a sequence representing the responsibilities of the robotic arm, and a sequence representing the positions of the support points.

[0115] Specifically, through deep reinforcement learning, the PPO algorithm is used to train the deep network to obtain the policy network of the space truss assembly sequence. Therefore, first, a deep network needs to be generated. The deep network includes:

[0116] Construct the state space S at time t t as:

[0117] S t = [s dis , s sup , s truss

[0118] where s dis is a scalar representing the duty assignment of the robotic arm. For the dual-arm task scenario, there are only two duty assignment methods. When the value is 1, it means that robotic arm 1 is the support arm and robotic arm 2 is the operating arm. When the value is 0, it means that robotic arm 2 is the support arm and robotic arm 1 is the operating arm. s sup is a scalar representing the support point position, corresponding to the ID number of the truss rod. When s sup = i, it means that the support point is at the centroid position of the truss rod with ID number i. s truss is a scalar obtained by dimensionality reduction using the assembly state S of the space truss structure T and represents the current state of the truss structure;

[0119] Construct the action space A at time t t as:

[0120] A t = [a1, a2, a3]

[0121] where a1 is a scalar representing the ID number of the truss component to be assembled, a2 is a scalar representing whether to perform the support arm replacement operation, and a3 is a scalar representing the support point position selection. When a2 = 0 and a3 = s sup , no support arm or support point replacement operation is performed. When a2 = 1, only the support arm replacement operation is performed without changing the support point position. When a2 = 0 and a3 ≠ s sup , only the support point position is changed without changing the support arm. When a2 = 1 and a3 ≠ s sup , both the support arm replacement operation and the support point position change are performed;

[0122] Construct the reward function R as:

[0123] R = -L

[0124] where L is the comprehensive cost evaluation index;

[0125] Build a deep network, which includes a policy network π θ and a value network Vφ , the policy network π θ takes the state space S t as the input and the action space A t as the output; the value network V φ takes the state space S t as the input and the value as the output.

[0126] Specifically, the policy network π θ and the value network V φ adopt the same network structure, where the input layer has 3 neurons; the hidden layer has 3 layers, and the number of neurons in each hidden layer gradually decreases, being 256, 128, and 64 neurons respectively. The ReLU function is used as the activation function, and the output layer has 202 neurons, using Softmax as the activation function.

[0127] Utilize a deep network to train the policy network for the assembly sequence of the space truss. The training steps are as follows:

[0128] Step 1: Obtain the connection matrix of the space truss structure, the initial poses of each truss component, the target poses of each truss component, the base poses of the two robotic arms, and the initial end poses;

[0129] Step 2: Let the current time step t = 1;

[0130] Step 3: Take the state space S t = [s dis , s sup , s truss at time step t as the input of the policy network π θ and obtain the action space A t = [a1, a2, a3] at time t;

[0131] Step 4: According to the assembly constraint representation, judge whether A t meets the robotic arm operation space constraint. If the judgment result is no, execute Step 3. If the judgment result is yes, then enter Step 5;

[0132] Step 5: Calculate the reward function value r t according to the reward function R;

[0133] Step 6: Combine S t , A t , r t , and S t+1 to construct an experience pool;

[0134] Step 7: Let t ← t + 1, and loop to execute Steps 3 to 7 until all truss components are traversed;

[0135] Step 8: Using the constructed experience pool, update the policy network π θ and the value network V φ parameters,;

[0136] Step 9: When all the data in the experience pool has been used for training, end the training and obtain the policy network π of the space truss assembly sequence θ .

[0137] According to the above method, for a double cube truss structure, as Figure 4 shown, in the figure, the numbers represent the ID numbers of the truss components, the black dots represent the connectors, and the straight lines represent the truss bars, perform the dual-arm coordinated assembly sequence planning to achieve experimental simulation. The rectangular truss structure is as Figure 2 shown, the rectangular truss structure includes 12 connectors and 20 truss bars, and the truss component set Q = [q1, q2,..., q 32 ; when i ≤ 4 or 9 ≤ i ≤ 12, the component mass m i = 1, the number of interfaces a i = 3; when 5 ≤ i ≤ 8, the component mass m i = 1, the number of interfaces a i = 4; when i ≥ 12, the component is a truss bar with a length of 0.2, the component mass m i = 1, the number of interfaces a i = 2.

[0138] The hyperparameter settings of the PPO algorithm are shown in Table 1.

[0139] Table 1, Hyperparameters of the PPO algorithm

[0140]

[0141] Select the ant colony algorithm for comparison. The optimized graph of the comprehensive cost after normalization and weighting is as Figure 5 shown. The coordinates in the figure represent the average value of the comprehensive cost of the trajectories generated by the PPO algorithm and the ant colony algorithm when iterating to the current number of times. From Figure 5 it can be seen that the ant colony algorithm converges earlier, but may fall into a local optimal solution and fail to find the optimal dual-arm coordinated assembly sequence. The PPO algorithm converges at about 420 times. The average value of the comprehensive cost of the dual-arm coordinated assembly sequence generated by the converged policy is 25.9, and the ant colony algorithm falls into a local optimal point at about 350 times, and the average value of the comprehensive cost of the dual-arm coordinated assembly sequence generated is 27.8. The comprehensive performance of the PPO algorithm in solving the assembly sequence of this truss structure is improved by 7.3% compared with the ant colony algorithm.

[0142] Specifically, finally obtain the space truss assembly sequence S e :

[0143] Se = [(13, 0, 13), (2, 0, 13), (18, 1, 18), (6, 1, 18), (21, 1, 13), (14, 1, 13), (22, 1, 13),

[0144] (1, 1, 22), (17, 1, 22), (3, 1, 13), (15, 1, 14), (7, 0, 14), (26, 1, 13), (0, 1, 14), (27, 1, 22),

[0145] (19, 0, 22), (16, 1, 18), (23, 1, 19), (4, 0, 21), (5, 0, 17), (10, 0, 14), (29, 0, 23),

[0146] (9, 1, 17), (12, 1, 17), (20, 0, 17), (25, 1, 17), (30, 1, 19), (28, 1, 20), (24, 1, 20),

[0147] (8, 1, 14), (11, 1, 30), (31, 1, 30)]

[0148] Spatial truss assembly sequence S e Consists of multiple tuples, each tuple corresponding to an action. The three elements of the tuple represent the spatial truss assembly order, the robotic arm duty assignment, and the support point position in sequence. The value of the first element in the tuple is the ID number of the truss component. The assembly order of the truss components with corresponding ID numbers is represented by the sequence of the tuples in the sequence. When the value of the second element in the tuple is 1, it means robotic arm 1 is the support arm and robotic arm 2 is the operating arm. When the value is 0, it means robotic arm 2 is the support arm and robotic arm 1 is the operating arm. The value of the third element in the tuple is the ID number of the truss component, indicating that the support point is on the truss component with this ID number.

[0149] Putting the first elements of each tuple in the assembly sequence together forms the spatial truss assembly order sequence S eseq :

[0150] S eseq = [13, 2, 18, 6, 21, 14, 22, 1, 17, 3, 15, 7, 26, 0, 27, 19, 16, 23, 4, 5, 10, 29,

[0151] 9, 12, 20, 25, 30, 28, 24, 8, 11, 31]

[0152] Putting the second elements of each tuple in the assembly sequence together forms the robotic arm duty assignment sequence S edis :

[0153] S edis = [0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 1, 1, 1, 0, 1, 1, 0, 0, 0, 0, 1, 1, 0, 1, 1, 1,

[0154] 1, 1, 1]

[0155] Putting together the third elements of each tuple in the assembly sequence, the formed sequence is the support point position sequence S esup :

[0156] S esup = [13, 13, 18, 18, 13, 13, 13, 22, 22, 13, 14, 14, 13, 14, 22, 22, 18, 19, 21,

[0157] 17, 14, 23, 17, 17, 17, 17, 19, 20, 20, 14, 30, 30].

[0158] The technical solution of the embodiment of the present invention has the following beneficial effects:

[0159] For the space truss assembly task, according to the energy consumption cost and time cost of the dual-arm coordinated support point replacement operation in the space truss assembly, a support point replacement operation cost function is generated, enabling the sequence planning method for space truss assembly to consider the support point replacement cost; then, based on the relationship between truss components, the assembly state, and the assembly constraints, a space truss assembly task characterization model is generated. On this basis, combined with the support point replacement operation cost function, a comprehensive cost evaluation index is designed, a reinforcement learning framework is generated, and through training with the PPO algorithm, a policy network for the space truss assembly sequence is obtained, realizing the solution of the space truss assembly sequence. The obtained sequence has better comprehensive performance compared with the sequence obtained by the traditional ant colony algorithm.

[0160] The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present invention shall be included within the scope of protection of the present invention.

[0161] The content not detailedly described in the specification of the present invention belongs to the well-known technology of those skilled in the art.

Claims

1. A sequential planning method for space truss assembly, characterized in that The method includes: For the spatial truss assembly task, according to the energy consumption cost and time cost of the dual-arm coordinated support point replacement operation in the spatial truss assembly, generate a support point replacement operation cost function; Generate a representation model of the spatial truss assembly task; According to the support point replacement operation cost function, assembly consumption time, manipulator disturbance torque, and manipulability of the manipulator, determine a comprehensive cost evaluation index; According to the comprehensive cost evaluation index and the representation model of the spatial truss assembly task, generate a deep network, and use the deep network to train the policy network of the spatial truss assembly sequence to obtain the trained policy network of the spatial truss assembly sequence. The policy network is used to obtain the spatial truss assembly sequence, and the spatial truss assembly sequence includes a sequence representing the assembly order of multiple truss components, a sequence representing the responsibilities of the manipulator, and a sequence representing the support point positions.

2. The method according to claim 1, wherein The step of, for the spatial truss assembly task, according to the energy consumption cost and time cost of the dual-arm coordinated support point replacement operation in the spatial truss assembly, generating a support point replacement operation cost function includes: The generated support point replacement operation cost function is: Z = P grab + P loose + M1 + M2 Among them, Z represents the total cost of the support point replacement operation, and P grab represents the time cost and energy consumption cost required for the support arm to grasp the assembled truss structure in the next state, and P loose represents the time cost and energy consumption cost required for the support arm to release the assembled truss structure in the current state, and P grab and P loose are affected by the truss shape and the end effector of the manipulator; M1 represents the time cost and energy consumption cost required for the support arm to move during the support point replacement operation, and M2 represents the time cost and energy consumption cost required for the manipulator arm to move during the support point replacement operation.

3. The method according to claim 1, wherein The method for generating a representation model of the spatial truss assembly task includes: Generate the self-attributes of the truss components; Generate the spatial truss structure connection matrix C; Generate an assembly state representation; Generate assembly constraint representations, including truss component connection constraints and truss component assembly order constraints; According to the self-attributes of the truss components, the spatial truss structure connection matrix C, the assembly state representation, and the assembly constraint representations, determine the representation model of the spatial truss assembly task.

4. The method according to claim 3, characterized in that, The self-attributes of the truss components include: Q = [q1, q2, …, q r ​ A i = [a i , m i ​ Among them, Q represents the set of truss components, and A i represents the self-attribute of the i-th truss component, and q i represents the i-th truss component. Suppose there are r truss components in total, including m connectors and h truss bars. When i ≤ m, q i represents a connector. When i > m, q i represents a truss bar, and a i represents the number of interfaces of the truss component q i , and a i ≥2, and m i represents the mass of the truss component q i .

5. The method according to claim 3, wherein The spatial truss structure connection matrix C: where c ij = c(i, j), and C represents the connection relationship between the truss members.

6. The method according to claim 3, characterized in that, The generated assembly state representation is: S T = [p1, p2,..., p r ​ Among them, W i represents the description of the assembly states of all adjacent truss components of the truss component q i . When a truss component has multiple adjacent truss components, they are sorted according to the original ID numbers of the adjacent truss components. w i1 ~w ia represent the assembly states of a adjacent truss components of the truss component q i . 0 indicates that the adjacent truss component does not exist, 1 indicates that the adjacent truss component is not assembled, 2 indicates that the adjacent truss component is assembled, and S T represents the assembly state of the space truss structure.

7. The method according to claim 3, wherein The truss component connection constraint is: K i = (1 - p i )h i where h i represents the number of adjacent truss members q i that have been assembled, and p i is consistent with the meaning of p defined by the assembly state characterization, and K i ≠ 0 indicates that the truss member q i satisfies this constraint condition; i ​ The truss component assembly order constraint is: H i = (w i1 - 1)(w i2 - 1), (i > m) Among them, w i1 With w i2 represents the truss q i The assembly state of two adjacent truss components, the assembly order constraint condition H of the spatial truss components i =0, indicating truss q i At least one of the adjacent truss parts is not assembled.

8. The method according to claim 1, characterized in that, The step of, according to the support point replacement operation cost function, assembly consumption time, manipulator disturbance torque, and manipulability of the manipulator, determining a comprehensive cost evaluation index includes: The determined comprehensive cost evaluation index L is: L = γ1ε1t all + γ2ε2τ + γ3ε3ω + Z Among them, γ1 represents the weight of the time cost index, γ2 represents the weight of the assembly stability index, and γ3 represents the weight of the operability index; ε1 represents the normalization parameter of the time cost, ε2 represents the normalization parameter of the assembly stability, and ε3 represents the normalization parameter of the operability; t all = t d + t o , t all represents the assembly consumption time, t d represents the running time of the path planning algorithm, t o is the time consumed by the robotic arm to perform the assembly task; τ = ||r × F||, τ represents the disturbing torque of the robotic arm, r represents the position vector from the rotation center to the force application point, and F represents the applied assembly force; ω represents the operability of the robotic arm, J(q) is the Jacobian matrix of the robotic arm when the joint angle is q; Z represents the support point replacement operation cost function.

9. The method according to claim 1, wherein The step of, according to the comprehensive cost evaluation index and the representation model of the spatial truss assembly task, generating a deep network includes: Construct the state space S at time t t as follows: S t = [s dis , s sup , s truss ​ Among them, s dis is a scalar representing the duty assignment of the robotic arm. For a dual-arm task scenario, there are only two duty assignment methods. When the value is 1, it means that robotic arm 1 is the support arm and robotic arm 2 is the operating arm. When the value is 0, it means that robotic arm 2 is the support arm and robotic arm 1 is the operating arm. s sup is a scalar representing the position of the support point, corresponding to the ID number of the truss rod. When s sup = i, it means that the support point is at the centroid position of the truss rod with the ID number i. s truss is a scalar obtained by dimensionality reduction using the assembly state S T of the space truss structure, representing the current state of the truss structure; Construct the action space A at time t t which is as follows: A t = [a1, a2, a3] Among them, a1 is a scalar representing the ID number of the truss component to be assembled, a2 is a scalar representing whether to perform the support arm replacement operation, and a3 is a scalar representing the selection of the support point position. When a2 = 0 and a3 = s sup no support arm or support point replacement operation is performed. When a2 = 1, only the support arm replacement operation is performed without changing the support point position. When a2 = 0 and a3 ≠ s sup the support arm is not replaced, and only the support point position is replaced. When a2 = 1 and a3 ≠ s sup both the support arm replacement operation and the support point position replacement operation are performed; Construct a reward function R as: R = -L where L is the comprehensive cost evaluation index; Construct a deep network, which includes a policy network π θ and a value network V φ , the policy network π θ takes the state space S t as input and the action space A t as output; the value network V φ takes the state space S t as input and the value as output.

10. The method according to claim 9, wherein The step of, using the deep network to train the policy network of the spatial truss assembly sequence to obtain the trained policy network of the spatial truss assembly sequence includes: Step 1: Obtain the spatial truss structure connection matrix, the initial poses of each truss component, the target poses of each truss component, the base poses of the two manipulators, and the initial end poses; Step 2: Let the current time step t = 1; Step 3: Take the state space S at time step t t = [s dis , s sup , s truss as the input to the policy network π θ and obtain the action space A t = [a1, a2, a3] at time t; Step 4: Determine A based on the assembly constraint representation t Check whether it meets the manipulator operating space constraint. If the judgment result is no, execute Step 3. If the judgment result is yes, proceed to Step 5; Step 5: Calculate the reward function value r according to the reward function R t ; Step 6: Combine S t , A t , r t , S t+1 to build an experience pool; Step 7: Let t ← t + 1, and loop through Steps 3 to 7 until all truss components are traversed; Step 8: Use the constructed experience pool and the PPO algorithm to update the parameters of the policy network π θ and the value network V φ ; Step 9: When all the data in the experience pool has been used for training, end the training and obtain the policy network π of the spatial truss assembly sequence θ .

Images