An evolutionary UAV landing structure design method based on the origami principle

Abstract

The present invention discloses a design method for an evolutionary UAV landing structure based on the origami principle, belonging to the technical field of UAVs. The landing structure of the UAV combines graph theory and the origami principle, calculates and conducts spatial search on the key points in the origami principle as elements in the graph, and constrains the entire design with the elements in the real physical scenario, enabling the design of the entire structure to be independent of human subjective will and allowing it to be automatically designed according to the motion requirements. The main part of the landing structure consists of a simple class capsule geometric body defined initially, and the joint part is represented by a simple cylinder for rotation and torsion. The connection method and force-bearing situation between the main part and the joint are optimized by applying combined physical parameter model predictive control to it in a physical real simulator and are obtained by screening through bistable conditions. Moreover, the finally generated physical model is easy to manufacture.

Description

Technical Field

[0001] The present invention discloses a design method for an evolutionary UAV landing structure based on the origami principle, belonging to the technical field of UAVs. Background Art

[0002] During the process of UAVs performing tasks such as transportation, detection, and rescue, generally, precision measuring instruments and valuable sensors are carried on the UAVs. When the UAV encounters unexpected situations such as strong winds, hailstorms and other bad weather, which causes the UAV to get out of control, it will deviate from the expected flight route, thus causing damage to the valuable items on the UAV. The current UAV landing devices on the market are mainly the above two structures. The first structure is a simple rigid landing gear. Although its structure is simple, it cannot be well applied to the landing of UAVs in many complex scenarios. For example, gentle slopes and potholed ground are likely to cause the overturning of the UAV body and damage to the wings. And because of its simple structure, it does not have a certain adjustment ability in case of emergencies, which brings certain difficulties to the task execution of the UAV. In addition to the difficulties during the task execution, in the later maintenance of UAVs, there are countless examples of the body being damaged due to impact and needing to be repaired or even scrapped in the fields of public UAVs and civilian UAVs every year, which greatly increases the maintenance cost in the UAV field. The second structure uses a servo motor to control and adjust the bottom support structure. Although it has a certain improvement in the UAV landing buffer compared with the first structure and can play a certain protective role in the landing of the UAV, each joint needs a servo motor as a support, which greatly increases the weight of the whole machine, reduces the loadable weight of the UAV, and limits the function during the transportation task. On the other hand, the control of this structure requires power supply for the servo motor. At present, most UAVs use lithium batteries for power supply. This structure undoubtedly becomes a burden on the normal operation of the UAV. In the context of the increasing miniaturization and lightweight of UAVs, the increase in the overall weight and the pressure of additional power supply will greatly reduce the endurance of the UAV, hindering the development and progress of UAV development. Summary of the Invention

[0003] The purpose of the present invention is to provide a design method for an evolutionary UAV landing structure based on the origami principle to solve the problem of poor buffering ability of the UAV landing structure in the prior art.

[0004] A design method for an evolutionary UAV landing structure based on the origami principle includes:

[0005] Step 1: Take the key stress points of the origami structure. The key stress points are the intersections of the creases, and the extracted key stress points are regarded as nodes and connections;

[0006] Step 2: Using graph grammar, assign geometric units with physical features in structure to key stress points, and use graph grammar to assign an attribute label to each key stress point. The attribute label determines the future structural connection sequence and joint type, and match the joint type with the main capsule structure according to the absolute distance between key stress points;

[0007] Step 3: The landing structure includes four support structures. A torsion joint is connected to one support structure, and a main capsule structure is connected behind the torsion joint. Another torsion joint is connected to the end of the main capsule structure, and a main capsule structure is connected behind the torsion joint. Another torsion joint is connected to the end of the main capsule structure... until the last key point with an attribute label is added to the support structure to form a complete support structure;

[0008] According to the principle of retaining the most important load-bearing members, simplify the adjacency matrix of key stress points. In the adjacency matrix, the points connected to the drone are expanded into threaded holes on the flat plate during the actual process;

[0009] Step 4: Use the geometric unit with graph information as the input of the search loop, and use the model predictive control in the actual landing physics engine to evaluate the performance to select candidate structures. For the search process, use the greedy algorithm to basically search the feasible space of the structure;

[0010] Step 5: Screen the candidate structures, and finally obtain the output of a drone landing structure model with bistable characteristics.

[0011] In the greedy algorithm, a heuristic function is introduced to make the search algorithm have a specific directionality. The reward function r(t) is defined as:

[0012]

[0013] In the formula, and both represent weights, indicating the directionality of the calculation model. and respectively represent the direction and speed required by the model. During the whole process, apply model predictive control to evaluate and optimize the action process of the generated structure descending. The optimization principle is to finally obtain the maximum reward value within the iteration process, and update the reward value as a reference for heuristic function training. The principle of heuristic function training is to sample the candidate structures searched, use the heuristic function value as the evaluation criterion, and use the squared loss function to update the heuristic function value.

[0014] The adjacency matrix includes:

[0015] Let G be a graph containing vertices v1, v2…v nFor the graph, the n×n adjacency matrix of G is defined as A, and the elements in A are denoted as [A] i,j :

[0016]

[0017] The calculation of the absolute distance between critical stress points includes:

[0018] Introduce an arbitrary positive integer k to determine the (i,j) term of A k and the relationship between vertex v i and v j When k = 1, [A] i,j = 1, that is, the number of times of advancing along an edge between vertex v i and v j For any k-edge advancing process, there exists an h such that this advancing process is considered as the combination of a (k - 1)-edge advance from v i to v h and an edge from v h to v j The total number of these k-edge advances is:

[0019] (The number of (k - 1)-edge advances from v i to v h );

[0020] According to the induction hypothesis, rewrite the total number of k-edge advances as:

[0021]

[0022] That is, for any positive integer k, the (i,j) term of A k is equal to the number of advances before k edges from v i to v j When the length of the main capsule structure is fixed, the absolute distance between critical stress points is obtained from the number of advances.

[0023] Model predictive control includes:

[0024] For the monitoring of structural actions, use model predictive control MPC, use the MPPI algorithm to generate optimized control inputs for each design, and MPC maintains a sliding window of control inputs for H time steps, representing the "best" control inputs found so far;

[0025] At the start of each iteration, K perturbation sequences of the control input are sampled according to the initial input sequence U, and each control input sequence is applied to a separate instance of the simulation. The new sequence U is calculated as a weighted sum of the perturbed sequences, where the weight of each sequence depends on the discounted sum of rewards of that simulation. The first control input of U is appended to the optimized control sequence as the output;

[0026] U is shifted forward by one time step in preparation for the next iteration, padding with zeros if necessary. Each MPC time step corresponds to multiple time steps in the simulation, with the ratio being the control interval, and the control input is repeated within the control interval.

[0027] The heuristic function includes:

[0028] A learning heuristic function V(g) is used to inform and accelerate the search of the design space. The function V(g) aims to output the highest achievable performance among all these complete designs. A deep learning-based approach is adopted, and a graph neural network is used to create the learnable heuristic function. The graph heuristic search algorithm works through a cross-design phase, an evaluation phase, and a learning phase. In the cross-design phase, candidate robots are sampled under the guidance of the heuristic function. In the evaluation phase, candidate robots are evaluated in the simulation. In the learning phase, the heuristic function is improved based on the simulation data.

[0029] Compared with the prior art, the present invention has the following beneficial effects: The physical elements that the drone will face during landing are incorporated into the design process, enabling the drone to convert the high vertical impact force during landing into lateral structural torque, greatly reducing the damage risk of the drone in the face of unexpected situations and being more suitable for the mission execution of most current drones. There are many drone-assisted landing structures on the market that use servos to control the execution of the landing structure. Such a servo-based landing structure not only requires additional power supply for the servo to drive, but also increases the overall weight. Since the batteries on drones generally have limited capacity, this landing structure will reduce the energy utilization rate. The designed landing structure incorporates the bistable characteristics of the origami structure into the design, automatically returning to a stable state during both the takeoff and landing processes of the drone to prepare for the next takeoff and landing; the entire takeoff and landing processes are energy-converted by the characteristics of the structure itself, without the need for additional energy input and power supply, enabling the energy of the drone to be more utilized in mission execution and greatly improving the energy utilization efficiency. Compared with servo control, this structure does not require additional actuating components, reducing the cost of structural application. Compared with traditional landing structures, this structure can be assembled only through 3D printing and simple springs, with low cost and a price difference not much from that of traditional landing structures. On the other hand, due to its good buffering performance, this structure reduces the difficulty of using and maintaining the drone and reduces the cost of drone maintenance. Detailed implementation manner

[0030] To make the objectives, technical solutions, and advantages of the present invention clearer, the technical solutions in the present invention will be clearly and completely described below. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments in the present invention without creative efforts shall fall within the protection scope of the present invention.

[0031] A design method for an evolved drone landing structure based on the origami principle, comprising:

[0032] Step 1: Take the key stress points of the origami structure. The key stress points are the intersections of the creases, and the extracted key stress points are regarded as nodes and connections;

[0033] Step 2: Use graph grammar to assign geometric units with physical characteristics on the structure to the key stress points, and use graph grammar to assign an attribute label to each key stress point. The attribute label determines the future structure connection sequence and joint type, and matches the joint type with the main body capsule structure according to the absolute distance between the key stress points;

[0034] Step 3: The landing structure includes four support structures. A torsion joint is connected to one support structure, and a main body capsule structure is connected behind the torsion joint. Another torsion joint is connected to the end of the main body capsule structure, and then another main body capsule structure is connected behind the torsion joint. Another torsion joint is connected to the end of the main body capsule structure... until the last key point with an attribute label is added to the support structure to form a complete support structure;

[0035] According to the principle of retaining the most important load-bearing members, the adjacency matrix of the key stress points is simplified. In the adjacency matrix, the points connected to the drone are expanded into threaded holes on the flat plate during the actual process;

[0036] Step 4: Use the geometric unit with graph information as the input of the search loop, and use the model predictive control in the actual landing physical engine to evaluate the performance to select the candidate structure. For the search process, use the greedy algorithm to basically search the feasible space of the structure;

[0037] Step 5: Screen the candidate structures, and finally obtain the output of a drone landing structure model with bistable characteristics.

[0038] In the greedy algorithm, a heuristic function is introduced to make the search algorithm have a specific directionality. The reward function r(t) is defined as:

[0039]

[0040] In the formula, and both represent weights, indicating the directionality of the calculation model. and represent the direction and speed required by the model respectively. During the whole process, model predictive control is applied to evaluate and optimize the action process of the generated structure falling downward. The optimization principle is to finally obtain the maximum reward value within the iteration process, and update the reward value as a reference for heuristic function training. The heuristic function training principle is to sample the candidate structures searched out, use the heuristic function value as the evaluation criterion, and use the squared loss function to update the heuristic function value.

[0041] The adjacency matrix includes:

[0042] Let G be a graph containing vertices v1, v2... v n The n×n order adjacency matrix of G is defined as A, and the elements in A are represented as [A] i,j :

[0043]

[0044] The calculation of the absolute distance between key stress points includes:

[0045] Introduce an arbitrary positive integer k to determine A k The (i, j) entry of i and vertex v j When k = 1, [A] i,j = 1, that is, the number of forward steps of an edge between vertex v i and v j For any k-edge forward process, there exists an h such that this forward process is considered a combination of a (k - 1)-edge forward from v i to v h and an edge from v h to v j The total number of these k-edge forward steps is:

[0046] (The number of (k - 1)-edge forward steps from v i to v h );

[0047] According to the induction hypothesis, rewrite the total number of k-edge forward steps as:

[0048]

[0049] That is, for any positive integer k, the (i, j) entry of A k is equal to the number of forward steps before k edges from v i to v j When the length of the main capsule structure is fixed, the absolute distance between key stress points is obtained from the number of forward steps.

[0050] Model predictive control includes:

[0051] For the monitoring of structural actions, use model predictive control MPC, use the MPPI algorithm to generate optimized control inputs for each design, MPC maintains a sliding window of control inputs for H time steps, representing the "best" control inputs found so far;

[0052] At the start of each iteration, sample K perturbation sequences of the control input according to the initial input sequence U, and apply each control input sequence to a separate instance of the simulation. The new sequence U is calculated as the weighted sum of the perturbed sequences, and the weight of each sequence depends on the discounted sum of rewards of that simulation. The first control input of U is appended to the optimized control sequence as the output;

[0053] U is shifted forward by one time step to prepare for the next iteration, padding with zeros if necessary. Each MPC time step corresponds to multiple time steps in the simulation, and the ratio is the control interval, and the control input is repeated within the control interval.

[0054] The heuristic functions include:

[0055] Use a learning heuristic function V(g) to inform and accelerate the search of the design space. The goal of the function V(g) is to output the highest achievable performance among all these complete designs. Adopt a deep learning-based method and use graph neural networks to create a learnable heuristic function. The graph heuristic search algorithm works through cross-design phase, evaluation phase, and learning phase. In the cross-design phase, sample candidate robots under the guidance of the heuristic function. In the evaluation phase, evaluate the candidate robots in simulation. In the learning phase, improve the heuristic function based on the simulation data.

[0056] Define a set of rules for the design of the structures expected in the research. The present invention applies this recursive graph grammar and defines it as a tuple:

[0057]

[0058] Where N and T are the sets of non-terminal and terminal symbols respectively, A is the set of attributes of some terminal symbols, rare production rules, and S ∈ N is the start symbol. Non-terminal symbols are temporary graphical nodes that help the application in the text construct different body and leg parts. Terminal symbols are the final graphical nodes that represent the physical components (such as links, joints, etc.) of the generated structure, and a graph consisting of only terminal symbols is called a complete robot design, while all other graphs are called partial robot designs. In addition, attributes are assigned to several terminal symbols. These attributes define the initial state of the robot by determining the initial relationships and angles between the structural components. Each production rule from R takes the following form:

[0059] Q → W;

[0060] Where Q ∈ N is a non-terminal symbol. And W is a graph with at least one node, whether it is a non-terminal symbol or a terminal symbol. When the grammar rule is applied to the current structure graph, W is replaced by detecting the occurrence of Q. Since the grammar is recursive, the non-terminal symbol Q can appear again on the right side of the rule as part of W. The recursive graph grammar allows indirect representation of various forms of structures, including many complex structural shapes.

[0061] The list of component-based rules in the framework and the overall structure that is ultimately used as a production reference. Due to the syntax-based method and the proposed rule set, the list of components for structure construction can be easily adjusted. There is no requirement to apply the structural rules before the component-based rules, and vice versa. The recursive grammar is designed in such a way that the number of legs and body segments can potentially be infinite. To limit the design space, a recursive counter is used in the algorithm, which calculates the total number of derivation steps. The maximum value was set to 40 in the experiment. Increasing this parameter will allow for the creation of more complex designs, but will also increase the running time of the algorithm. The MPPI implementation outline is given in Algorithm 1.

[0062] Table 1 Model Predictive Control Based on MPPI

[0063]

[0064] In the heuristic search algorithm, the three phases are repeated on N segments, or until they reach an optimal design, as described in Algorithm 2, splitting Table 2 into multiple tables.

[0065] Table 2.1 Heuristic Graph Search in Bistable Space

[0066]

[0067] Table 2.2 Heuristic Graph Search in Bistable Space

[0068]

[0069] Table 2.3 Heuristic Graph Search in Bistable Space

[0070]

[0071] The present invention follows the directed acyclic graph form of recursive graph grammar and generates a spatial structure according to the graph grammar rules. The graph grammar can be represented as: G = {V B , V L , E}, where the vertices {v o} ∈ V B correspond to body parts; the vertices {v i , v j , v k} ∈ V L correspond to the vertices of each horizontal plane in the steady state space; the edges {e(i,j), e(j,k)} ∈ e represent the connections between adjacent two planes; the arrow direction of the corresponding edges represents the expansion direction of the structure in the graph grammar; V B , V L and E are restricted in the bistable space B defined above.

[0072] For comparison, the search effectiveness of Monte Carlo Tree Search (MCTS), especially the UCT algorithm, was also referred to. Algorithm 3 gives an output line of the MCTS implementation. The state of the algorithm is represented by a directed acyclic graph of nodes, which represents partial designs and related statistics. The directed edges between nodes represent operations or rule applications. MCTS implementations tend to differ in the stored statistics. Since the problem setting is single-shot and non-aggressive, the stored visit count and the maximum reward are chosen. Similar to heuristic graph search, each iteration of the MCTS implementation consists of design, evaluation, and learning phases. The design phase combines the selection and expansion steps of a standard MCTS iteration. The selection step starts from the root of the tree, representing the start symbol of the grammar. The edges are repeatedly followed until a node without children (a leaf node) is reached. When multiple edges are available, the edge with the highest UCT score is selected:

[0073]

[0074] where Q s,a (t) is the maximum result of all iterations in which the rule is applied to the graph, N s (t) is the number of visits to s, and N s,a (t) is the number of iterations in which the rule is applied to s. If the last reached node represents a partial design, an expansion step is performed: a randomly selected rule is applied, and the node of the resulting design is added to the tree. Without further modifying the tree, rules are randomly selected and applied until a complete design is obtained. The complete design is simulated and evaluated using MPC in a way similar to graph heuristic search. In the learning phase, the statistics of the nodes in the search tree that were visited during the design phase are updated. The visit count of each visited node is incremented, and if its maximum reward is higher, the evaluation result is substituted. These updated statistics will guide the node selection in the next iteration.

[0075] Table 3 Heuristic Graph Search in Bistable Space

[0076]

[0077] The present invention implements a graph heuristic search algorithm in Python and implements simulation and MPC in C++. The experiments were carried out on a graphics card with an i7 CPU and two RTX 3090s. Each iteration of GHS takes less than 1 second for each possible candidate structure in the design phase, 40 - 60 seconds in the MPC evaluation phase, and 6 - 8 seconds in the learning phase. Since each possible candidate structure is sampled independently, the design phase can be fully parallelized to further accelerate. The time bottleneck in the MPC evaluation phase indicates the necessity of the graph heuristic search algorithm, which can find the best-performing landing structure when evaluating a much smaller number of structural designs. The parameters applied in the experiments are listed in Tables 4 and 5 respectively.

[0078] Table 4 Parameters in heuristic graph search

[0079]

[0080] Table 5 Parameters in MPC and simulation

[0081]

[0082] To make the designed structure function in practice, the principle of equivalent substitution is defined, mainly for those components that play a decisive role in the structural movement, such as rigid connections, joints, and torques. Here, structures that are as centrally symmetric as possible in the design are equivalently substituted to maintain the center of gravity of the drone during flight. A single rolling joint rotating along the axis is equated to a rotary joint, and the rotary joint is replaced by a thread for the torsional joint. The distance of the thread is set according to the iterative change of the model movement. To adjust the joint torque to the required value, elastic elements are embedded at each joint so that the entire structure meets the design requirements. The green cylinder represents the torsional joint in the simulation, and the orange cylinder represents the rolling joint. A flat plate equivalently replaces the body for installation. The position of the holes on the flat plate for supporting the structure is obtained by equivalently scaling the node parameters of the body of the simulated structure by approximately 1:3, which is the same as the ratio of the supporting structure. The prefabricated structure consists of three main parts: a connecting plate (corresponding to the first geometric plane of the proposed origami structure), an action actuator (corresponding to the valley folds in the origami structure), and a steering component (corresponding to the turning points in the origami structure).

[0083] In addition, the actual structure of the customized proportional landing system and the components of the key parts were made, and the tension return spring was stretched to ensure the consistent movement of the joints during landing. The stainless-steel shaft and torsion spring are equivalent to a rolling joint. The 3D-printed mating parts and the threads connected by the return spring in the middle are used to form a torsional joint. All the basic parts are obtained by 3D printing using polylactic acid (PLA) material. The springs are ordered with the calculated joint torques. Except for the different structural forms, the remaining components of the actual structure of the baseline algorithm are the same as those of the bistable structure design.

[0084] To further optimize the equivalent replacement and analyze the structural performance after the equivalent replacement, a force sensor was used to conduct a dynamic analysis on some instantiated structures. The lifting platform below moved upward at a constant speed, and the force sensor above output the force when the structure deformed, and sampled the force value at a frequency of 50 Hz to obtain a force-displacement image.

[0085] To further illustrate the performance of the designed structure and its protective ability for the UAV landing compared with other structures, experiments on different structures were added in multiple outdoor scenarios. Conducting landing experiments on different structures under external disturbances and with only general-precision positioning can better highlight the advantages of the design and its protective ability to improve UAV landing.

[0086] Different from the UAV architecture in the indoor experiment, the outdoor UAV only requires GPS positioning information and does not need an optical motion capture system for positioning and controlling the UAV, so there is no need for a microcomputer equipped with ROS. In the outdoor UAV architecture, the Intel NUC was removed, and GPS was used to provide positioning information for the UAV. The remaining components are basically the same as those in the indoor architecture.

[0087] For the experimental process, various structures were experimented and verified multiple times in different terrains. For the outdoor UAV architecture, to ensure the unity of experimental variables, the weights of the systems equipped with different landing structures were exactly the same. The weights of the bistable structure and the baseline structure were exactly the same, at 383 g. The typical structure added equal-weight counterweights for controlling variables. The weather conditions in the experimental environment on the day were basically the same: wind speed of level 3-4, temperature of 10 degrees Celsius, and no serious signal interference causing GPS failure.

[0088] Based on the results of multiple experiments on UAV systems with different landing structures outdoors, it can be seen that for UAV systems equipped with typical landing structures (i.e., without additional auxiliary equipment), although they do not require additional actions during landing, they will exhibit relatively large peak accelerations whether on flat ground, grassland, or slopes, which is not beneficial for many devices. Additionally, when landing on cluttered terrains such as grasslands, it is difficult for typical landing structures to avoid the influence of dead grass, fallen leaves, etc. on the ground. For the landing structure designed by the baseline algorithm, although the peak acceleration has decreased, due to the limitations of the landing structure during the landing process, there will be obvious jitters at the moment of landing, along with a position error of approximately 10 cm, which is not conducive to installing high-precision observation equipment on the UAV. However, the bistable structure designed based on the origami structure and graph search method not only achieves good buffering performance but also, due to its bistable nature, can control the position error within a range of 2 cm, making it more conducive to task execution and perching on complex ground surfaces.

[0089] The present invention designs and manufactures components according to the model generated by the design rules described above. Through the corresponding force and rotation relationships, the structures at the corresponding parts of the feature points in the model are equivalently reconstructed in a real-world manufacturable model, and the reconstructed model is further analyzed and verified. Actual tests are carried out in indoor and outdoor environments. Through precise data comparison and macroscopic behavior performance, the design method combining graph search and origami principle is summarized and analyzed, and the experimental data is further processed. Both the analyzed experimental data and the actual performance of the UAV landing structure illustrate the superior performance of the method proposed in the present invention.

[0090] 1) Based on the force and motion characteristics of the feature points, the components in the model calculated in the previous chapter are equivalently replaced, and the composition method of its mechanical structure is further optimized. While meeting the defined landing method in the model, the mechanical components in the structure make it more robust and adaptable.

[0091] 2) The high-speed dynamic force value measurement system is applied to conduct secondary verification and mechanical property analysis on the 3D printed model. The reaction force of the structure during compression is transmitted to the data processing unit by the force sensing element, and the data is recorded through the software in the computer. Fitting the force and displacement curves of multiple experiments shows that this equivalent replacement method fully retains the bistable performance of the Kresling origami structure, enhancing the reliability when applied to UAV landings.

[0092] 3) By mounting a drone control framework under an optical motion capture system and utilizing the external positioning of the drone provided by the motion capture system, more accurate motion data of the drone system during landing is obtained; the above data is analyzed and verified using the finite difference method, proving that the data obtained is very accurate, further highlighting the performance difference between the structure designed in the present invention and other structures.

[0093] 4) A GPS is applied to implement an outdoor drone positioning control framework. Since GPS does not have high-precision positioning ability compared to the motion capture system, the positioning error of the drone is relatively large whether it is flying or landing. This further provides a complex environment for the application of this landing structure in practice, so as to further test the outdoor ability of this structure. After conducting multiple experiments on the drone under GPS navigation outdoors, the structure with bistable performance shows a better landing effect than other design schemes.

[0094] The above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements on some or all of the technical features, and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A design method for the landing structure of an evolutionary unmanned aerial vehicle based on the origami principle, characterized in that, Including: Step 1: Obtain the key stress points of the origami structure. The key stress points are the intersections of the creases. The extracted key stress points are regarded as nodes and connections; Step 2: Use graph grammar to assign geometric units with physical characteristics on the structure to the key stress points. Use graph grammar to assign an attribute label to each key stress point. The attribute label determines the future structure connection order and joint type. Match the joint type with the main body capsule structure according to the absolute distance between the key stress points; Step 3: The landing structure includes four support structures. A torsion joint is connected to one support structure, and a main body capsule structure is connected behind the torsion joint. Another torsion joint is connected to the end of the main body capsule structure, and a main body capsule structure is connected behind the torsion joint. Another torsion joint is connected to the end of the main body capsule structure, and a main body capsule structure is connected behind the torsion joint. Repeat connecting the torsion joint and the main body capsule structure until the last key point with an attribute label is added to the support structure to form a complete support structure; According to the principle of retaining the most important load-bearing members, simplify the adjacency matrix of the key stress points. In the adjacency matrix, the points connected to the drone are expanded into threaded holes on the flat plate during the actual process; Step 4: Use the geometric unit with graph information as the input of the search loop, and use the model predictive control in the actual landing physics engine to evaluate the performance to select the candidate structure. For the search process, use the greedy algorithm to perform a basic search on the feasible space of the structure; Step 5: Screen the candidate structures, and finally obtain an output of a drone landing structure model with bistable characteristics.

2. The design method of the evolutionary UAV landing structure based on the origami principle according to claim 1, characterized in that, In the said greedy algorithm, a heuristic function is introduced to make the search algorithm have a specific directionality. The reward function r(t) is defined as: In the formula, and both represent weights, indicating the directionality of the calculation model. and respectively represent the direction and speed required by the model. During the whole process, model predictive control is applied to evaluate and optimize the action process of the generated structure descending. The optimization principle is to finally obtain the maximum reward value within the iteration process, and update the reward value as a reference for the training of the heuristic function. The training principle of the heuristic function is to sample the searched candidate structures, use the heuristic function value as the evaluation criterion, and use the squared loss function to update the heuristic function value.

3. The design method of the evolutionary UAV landing structure based on the origami principle according to claim 2, characterized in that, The adjacency matrix includes: Let \(G\) be a graph containing vertices \(v_1, v_2, \ldots, v\) n and the \(n\times n\) adjacency matrix of \(G\) is defined as \(A\), and the elements in \(A\) are denoted as \([A]\) i,j :\ 4. The design method of the evolutionary UAV landing structure based on the origami principle according to claim 3, characterized in that, The calculation of the absolute distance between the key stress points includes: Introduce an arbitrary positive integer k to determine A k The (i, j) entry of i and the relationship between vertices v j When k = 1, [A] i,j = 1, that is, the number of forward steps of an edge between vertices v i and v j For any k-edge forward process, there exists an h such that this forward process is considered to be a combination of a (k - 1)-edge forward from v i to v h and an edge from v h to v j The total number of these k-edge forward steps is equivalent to the number of (k - 1)-edge forward steps from v i to v h ; According to the inductive hypothesis, rewrite the total number of k-edge advances as: That is, for any positive integer k, the (i, j)-th entry of A k is equal to the number of forward steps before k edges from v i to v j . When the length of the main capsule structure is fixed, the absolute distance between the key stress points can be obtained from the number of forward steps.

5. The design method of the evolutionary UAV landing structure based on the origami principle according to claim 4, characterized in that, Model predictive control includes: For the monitoring of the structure actions, use model predictive control MPC. Use the MPPI algorithm to generate optimized control inputs for each design. MPC maintains a sliding window of control inputs for H time steps, representing the "best" control inputs found so far; At the beginning of each iteration, sample K perturbation sequences of the control input according to the initial input sequence U, and apply each control input sequence to a separate instance of the simulation. The new sequence U is calculated as the weighted sum of the perturbed sequences. The weight of each sequence depends on the discounted sum of the rewards of that simulation. The first control input of U is appended to the optimized control sequence as the output; U is shifted forward by one time step to prepare for the next iteration. Fill zeros when the length of the control input sequence is insufficient. Each MPC time step corresponds to multiple time steps in the simulation, and the ratio is the control interval. The control input is repeated within the control interval.

6. The design method of the evolutionary UAV landing structure based on the origami principle according to claim 5, characterized in that, The heuristic function includes: A learning heuristic function V(g) is used to inform and accelerate the search of the design space. The goal of the function V(g) is to output the highest achievable performance among all these complete designs. A deep learning-based method is adopted, and a graph neural network is used to create a learnable heuristic function. The graph heuristic search algorithm works through a cross-design stage, an evaluation stage, and a learning stage. The cross-design stage samples candidate robots under the guidance of the heuristic function. The evaluation stage evaluates the candidate robots in simulation. The learning stage improves the heuristic function based on the simulation data.

Images