Asteroid intelligent cooperative detection planning method based on visual uncertainty feedback

Abstract

The invention discloses an asteroid intelligent cooperative detection planning method based on visual uncertainty feedback. The asteroid intelligent cooperative detection planning method comprises the steps of performing multi-detector asteroid detection dynamics modeling and coordinate system conversion relation modeling; constructing a multi-detector cooperative asteroid body detection scene; constructing an asteroid concept envelope model and performing grid division; defining detection visual uncertainty, environment uncertainty and information gain; and performing multi-detector intelligent cooperative detection decision planning based on visual uncertainty feedback. According to the method, the problem that the rough shape of a target asteroid must be achieved before detection in an existing mainstream scheme is solved, and the detection decision planning generalization ability of an algorithm on a deep space unknown target is improved; a planning decision is directly fed back by considering cloud computing environment gain of a body detection sampling point, so that the detection efficiency is improved and the energy consumption is reduced; according to the method, end-to-end feedback of the asteroid three-dimensional reconstruction detection effect on detector decision planning is realized, and the cost-effectiveness ratio of multiple detectors to asteroid detection is improved.

Description

Technical Field

[0001] The present invention relates to a multi-spacecraft asteroid exploration planning method based on visual uncertainty feedback, belonging to the technical field of multi-spacecraft collaborative exploration mapping. Background Art

[0002] NASA in the United States has conducted a large number of studies on the exploration of asteroid Bennu. It mainly rasterizes the surface polygons of the 3D model of asteroid Bennu, and feeds back the asteroid observation quality through measuring the incident angle, exit angle, sun-asteroid-spacecraft angle of each polygon by sunlight, and reward indicators such as obstacle avoidance between multiple detectors, so as to design integer programming or other optimization and reinforcement learning algorithms for decision-making planning of single-star or multi-detector collaborative asteroid exploration. This method only extracts the main influencing factors affecting the observation quality of the detector for the asteroid, and does not directly use the real-time feedback of the final task effect of the exploration for decision-making planning, which makes it difficult to ensure the final exploration effect and decision-making planning efficiency well. Moreover, in the initial design stage, a rough 3D model of the asteroid is required first for polygon rasterization, resulting in limited adaptability of the algorithm to unknown deep space environments. Summary of the Invention

[0003] The technical problem to be solved by the present invention is: overcoming the deficiencies of the prior art, providing an asteroid intelligent collaborative exploration planning method with visual uncertainty feedback, realizing the end-to-end feedback of the 3D reconstruction exploration effect of the asteroid on the detector decision-making planning, and improving the cost-effectiveness of multi-detector asteroid exploration.

[0004] The technical solution of the present invention is:

[0005] An asteroid intelligent collaborative exploration planning method with visual uncertainty feedback, comprising the following steps:

[0006] (1) Conduct dynamic modeling of multi-detector asteroid exploration and modeling of coordinate transformation relationships;

[0007] (2) Construct a multi-detector collaborative asteroid embodied exploration scenario;

[0008] (3) Construct an asteroid conceptual envelope model and grid division;

[0009] (4) Define detection visual uncertainty, environmental uncertainty and information gain;

[0010] (5) Conduct multi-detector intelligent collaborative exploration decision-making planning based on visual uncertainty feedback.

[0011] Further, the dynamic modeling of multi-detector asteroid exploration is specifically:

[0012] Construct a relative orbital coordinate system o-x with the centroid of the asteroid as the origin w y w z w and construct the asteroid body coordinate system o-x b y b z b , construct the multi-satellite body coordinate system o-x s1 y s1 z s1 、o-x s2 y s2 z s2 、o-x s3 y s3 z s3 、…;

[0013] Assume the detector as a particle and satisfy the relative orbital dynamics equation as follows:

[0014]

[0015] where, r represents the orbital radius; θ is the orbital rotation angle of the relative orbital coordinate system relative to the space inertial coordinate system; x, y, z respectively represent the three-axis coordinates of the satellite in the relative orbital coordinate system, and are the corresponding velocity and acceleration; f x , f y , f z is the acceleration of each axis of the detector in the relative orbital coordinate system, and its value is the sum of the perturbation acceleration and its own power; G x 、G y 、G z are the three-axis projections of the asteroid gravitational field in the relative orbital coordinate system; μ is the gravitational constant of the two-body system.

[0016] Furthermore,

[0017] The definition of the relative orbital coordinate system o-x w y w z w is: with the centroid of the asteroid as the origin, the direction from the centroid of the sun to the centroid of the asteroid is the x w axis, the direction of the asteroid's running speed is the y w axis, the z w axis is the direction of the asteroid's angular momentum, and it conforms to the right-hand rule with the x w axis and the y w axis;

[0018] The definition of the asteroid body coordinate system o-x b y b z b is: the x b axis is the direction of the asteroid's maximum inertia axis, zb The axis is in the direction of the small and large inertia axes of the asteroid, and y b axis is perpendicular to the x b axis and the y b axis satisfies the right-hand rule;

[0019] The multi-satellite body coordinate system o-x s1 y s1 z s1 、o-x s2 y s2 z s2 、o-x s3 y s3 z s3 is defined in the same way as the asteroid body coordinate system.

[0020] Furthermore, the establishment of the coordinate system transformation relationship specifically includes:

[0021] ① Transformation between the asteroid body coordinate system and the relative orbital coordinate system

[0022] Considering the rotation of the asteroid, a transformation relationship between the asteroid body coordinate system o-x b y b z b and the relative orbital coordinate system o-x w y w z w is constructed. Assuming that the zb axis of the asteroid body coordinate system overlaps with the z w axis of the relative orbital coordinate system, the relationship between the asteroid body coordinate system and the relative orbital coordinate system is as follows:

[0023]

[0024] where η = η0 + ωt, η is the angle between the x axis of the asteroid body coordinate system and the x axis of the relative orbital coordinate system, η0 is the initial value of this angle, ω is the rotational angular velocity of the asteroid body coordinate system relative to the orbital coordinate system, and t is the rotation time;

[0025] ② Transformation between the detector body coordinate system and the relative orbital coordinate system

[0026]

[0027] where α, β, γ are the angles corresponding to the coincidence of the relative orbital coordinate system with the detector body coordinate system after rotating around its own coordinate system's three axes; x si 、y yi 、z zi are the point coordinates in the detector i body coordinate system; x w 、y w 、z w are the point coordinates in the relative orbital coordinate system;

[0028] ③ Transformation between the detector body coordinate system and the asteroid body coordinate system

[0029] Among them, First, convert the asteroid body coordinate system to the relative orbital coordinate system, and then convert it to the detector body coordinate system.

[0030] Furthermore, the construction of the multi-detector collaborative asteroid embodied detection scenario is specifically as follows:

[0031] Construct a platform for the embodied detection scenario with the GAZEBO physics engine, calibrate the body coordinate systems of the asteroid and the detector body models and implant them into this physics engine, and at the same time set the conversion relationship between the coordinate systems; on this basis, implant the embodied model of the depth camera, calibrate the camera coordinate system and complete the conversion between the camera coordinate system and the detector body coordinate system.

[0032] Furthermore, the construction of the asteroid conceptual envelope model and grid division is specifically as follows:

[0033] Envelope the asteroid in a spherical envelope body, the radius of which is the maximum radius of the asteroid, called the asteroid conceptual envelope body, that is, the asteroid conceptual envelope model;

[0034] Grid division of the envelope body: First, design that the asteroid conceptual envelope body is composed of twenty spherical equilateral triangles, and then divide each spherical equilateral triangle into 4 small spherical equilateral triangles, and continuously divide the spherical envelope body based on this step;

[0035] The size of the elementary granularity of the equilateral triangle that makes up the asteroid conceptual envelope body depends on the field of view radius determined by the flight altitude and field of view angle of the detector, and the diameter of the triangular elementary body of the envelope body needs to be less than the field of view diameter of the detector; all the divided grids are the set U.

[0036] Furthermore, the definition of the detection environment uncertainty and information gain specifically includes:

[0037] (4.1) Define the visual uncertainty within the detection field of view

[0038] In the embodied detection environment, the spacecraft uses the depth camera to detect and sample, and obtains the three-dimensional feature points on the surface of the asteroid within the field of view, that is, the point cloud. Perform state probability estimation on the positions marked by the point cloud, and take the proportion of the number of point cloud feature points falling in each grid of the envelope body within the field of view to the total number of point clouds in this sampling as the visual uncertainty probability P of the grid (m,n) υ ;

[0039] Model the sensitivity distribution of the depth camera detection and perform a weighted sum with the probability obtained from the preliminary calculation of the point cloud to obtain a more reasonable probability distribution of the visual uncertainty of the asteroid surface grid

[0040]

[0041] Among them, M represents the sensitivity distribution function within the detection field of view of the detector; L(x) = exp(x) / (1 + exp(x)); D is the boundary of the field of view. When approaching the upper boundary, that is, the outermost edge of the field of view of D, the sensitivity is 0; is the tolerance, reflecting the most sensitive range within the field of view; ζ1 and ζ2 are coefficients;

[0042] (4.2) Define the uncertainty of the asteroid surface environment

[0043] Assign the probability of the asteroid environment uncertainty to each grid (m, n) ∈ U in the grid set U of the asteroid conceptual envelope. During the detection process of the detector, update the environmental uncertainty probability of each grid using the visual uncertainty within the detection field of view of the detector. When the environmental uncertainty of the entire grid set U of the envelope reaches the lowest, the purpose of this detection is achieved; represent the environmental uncertainty state of each grid on the asteroid surface during the spacecraft detection process in three states: occupied, unknown, and idle;

[0044] In the initial state, the asteroid surface environment is completely unknown, and the environmental uncertainty probability P occ (m, n) of all grids is set to 0.5, (m, n) ∈ U; when the probability reaches the set upper boundary P occ_threshold , the grid is considered occupied and the route to reach this point is no longer planned repeatedly; when P occ (m, n) is lower than the lower boundary P free_threshold , it is idle, that is, try to plan the route to detect here as much as possible; for the area not detected by the detector, it is marked as unknown, and P occ (m, n) remains 0.5;

[0045] The environmental uncertainty P mn (t) of the asteroid surface grid (m, n) is updated using the Bayesian probability update mechanism:

[0046] ① When the detector does not detect the grid (m, n)

[0047] P occ ((m, n), t) = P occ ((m, n), t - 1)

[0048] ② When the detector detects the grid (m, n)

[0049]

[0050] Among them, P d respectively represents the detector detection probability, that is, the visual uncertainty; P f represents the false alarm rate of faults; t represents the number of detector detection steps;

[0051] (4.3) Define the visual detection gain within the detection field of view

[0052] Suppose the detector detects the grid (m, n) on the asteroid surface. Due to the detector being at a certain height, all adjacent areas within the depth visual camera's detection field of view except the grid (m, n) will be observed, and the environmental grids outside the detection field of view will not be detected;

[0053] The environmental detection gain of the grid (m, n) is:

[0054] E(m, n) = -P occ (m, n) log P occ (m, n) - (1 - P occ (m, n)) log(1 - P occ (m, n))

[0055] The visual detection gain of adjacent grids within its detection field of view is also calculated according to this formula. For areas that have not been detected at all during the entire mission process, the detection information gain is judged to be the largest by this formula, and thus the spacecraft is guided to detect the undetected areas.

[0056] Furthermore, the multi-detector intelligent collaborative detection decision-making and planning based on visual uncertainty feedback is specifically as follows:

[0057] (5.1) Conduct information sharing

[0058] During the detection process, each detector is initialized with the visual detection gain information of all grids on the asteroid surface, that is, the environmental detection gain map of the asteroid surface. During the detection process, each detector will share information with each other and continuously update this map according to the visual uncertainty, environmental uncertainty within the detection field of view, and the visual detection gain calculation formula, so as to dynamically guide the detector to the grid with high detection gain;

[0059] (5.2) Conduct multi-detector collaborative asteroid detection decision-making and planning based on multi-agent reinforcement learning, that is, design an interaction training between the multi-agent reinforcement learning algorithm and the asteroid detection environment of the multi-detection instruments to generate a multi-spacecraft asteroid detection intelligent decision-making and planning model. The action space, state space, and reward function are designed as follows;

[0060] ① Action space

[0061] Design the action space as the three-axis accelerations [a of each detectorx , a y , a z , where a x , a y , a z are all less than the maximum acceleration a max ;

[0062] ② State space

[0063] [p i , v i , Δp ij , Δv ij , …], i = 1, 2, …, j ≠ i, where p i is the position of detector i, v i is the velocity of detector i, Δp ij is the position difference between detector i and detector j, and Δv ij is the velocity difference between detector i and detector j;

[0064] ③ Reward function

[0065] A. Collision reward

[0066] In the embodied detection environment, if a detector collides with an asteroid or detectors collide with each other during the training process, the embodied detection environment gives negative feedback R1 and ends the current training;

[0067]

[0068] Among them, R1 represents the collision reward between the detector and the asteroid. When the detector is greater than the radius d of the asteroid w and less than a certain flight range d w + Δ s , then a positive reward μ is given; Δ s is a certain flight height range of the detector from the surface of the asteroid, and p o represents the origin position in the body coordinate system of the asteroid, and p i represents the position of the detector in the body coordinate system of the asteroid;

[0069] When the detector is less than the radius d of the asteroid w , a negative reward -κ is given for the collision between the detector and the asteroid;

[0070] When the flight range of the detector exceeds d w + Δ s , a negative guidance reward -τ·|p i - p o | is given to pull it back to the set detection range; τ is a coefficient;

[0071]

[0072] Among them, R2 is the collision reward between detectors, and k and σ are both positive coefficients. When the detectors i and j are greater than each other's minimum safety threshold d th and less than the safety margin Δ ij , a positive reward is given to maintain this distance as much as possible; when the detectors i and j are less than each other's minimum safety threshold d th , a negative reward is given, and the closer the approaching distance is, the more severe the negative reward is; p i and p j represent the positions of the detectors i and j in the asteroid body coordinate system;

[0073] B. Guidance Reward

[0074] The detector tries to detect the direction with the maximum information gain within the field of view as the guidance mechanism;

[0075] R3 = λ·|p i -Z max_i |

[0076] Among them, λ is a coefficient, p i represents the position corresponding to the i-th detector, and Z max_i represents the position with the maximum information gain within the field of view detected at the position where the i-th detector is located; R3 is the guidance reward for the detector to move towards the direction with the maximum information gain within the field of view;

[0077] C. 3D Reconstruction Effect Reward

[0078] During the detection process, the uniformity of all grid sampling point clouds on the asteroid surface can most significantly indicate the asteroid detection and reconstruction effect. The uniformity and quantity of the depth camera sampling point clouds are represented by the standard deviation of the full-grid gain. The specific reward function is set as follows:

[0079]

[0080] Among them, num is the number of grids divided on the asteroid surface; S is the standard deviation; is a coefficient, and R4 is the 3D reconstruction effect reward, that is, the higher the standard deviation, the more uneven the overall point cloud coverage and collection on the asteroid surface, and the lower the reward; is the sum of the gains of each grid divided by the number of grids;

[0081] The total reward function is as follows:

[0082] R all = ρ1·R1 + ρ2·R2 + ρ3·R3 + ρ4·R4

[0083] Among them, ρ1, ρ2, ρ3, and ρ4 are coefficients, and R1, R2, R3, and R4 respectively correspond to the above reward functions;

[0084] ④ Training and testing

[0085] Initialize the states of multiple detectors and the parameters of the decision network. The multiple detectors update their states by interacting with the task environment, and update the parameters of the decision network according to the state gradient descent of multi-agent reinforcement learning. When the reward function of the multiple detectors stabilizes within a certain range and no longer increases, stop the training and save the model as the collaborative asteroid detection decision-making and planning model for multiple detectors.

[0086] Furthermore, test the trained collaborative asteroid detection decision-making and planning model for multiple detectors: instantiate the multi-agent reinforcement learning network and import the trained decision-making and planning model, initialize the states of multiple detectors and the environment. The multiple detectors obtain the environmental observation information through their respective local perceptions as the input of the decision network, output their respective accelerations, and update the detection plan through continuous iteration. Finally, end the test with the maximum environmental information gain no longer increasing as the algorithm evaluation criterion.

[0087] The beneficial effects brought by the present invention compared with the prior art are as follows:

[0088] The present invention relates to an asteroid intelligent collaborative exploration and planning method with visual uncertainty feedback. Aiming at the problem that traditional asteroid detection requires a rough three-dimensional model in advance, which is not conducive to the detection of unknown targets, a method for calculating the sampling probability of point clouds of an asteroid concept spherical grid envelope combined with an embodied detection depth camera is proposed, which can effectively solve the problem of insufficient generalization ability of the current multi-detector asteroid detection method for unknown asteroid targets. Aiming at the problem that the sparsity of depth camera point cloud acquisition easily leads to difficult grid probability calculation, a grid visual uncertainty probability distribution function weighted by the detection sensitivity probability distribution function is proposed, which can effectively solve the problem of difficult solution of the visual uncertainty on the surface of the detection target caused by sparse point cloud sampling within the detection field of view. Aiming at the problem of three-dimensional reconstruction detection decision evaluation of multiple spacecraft for asteroids, an asteroid surface full-grid gain standard deviation is proposed as the reward function, which can ensure the execution efficiency of the multi-spacecraft detection decision-making and planning method trained based on multi-agent reinforcement learning and the balance of the overall three-dimensional reconstruction effect.

[0089] Therefore, in summary, the present invention has at least the following advantages:

[0090] (1) Solve the problem that a rough shape of the target asteroid must be available before detection in the existing mainstream solutions, and improve the generalization ability of the algorithm for detecting and making decisions on deep-space unknown targets;

[0091] (2) Consider the direct feedback of the environmental gain of the embodied detection sampling point cloud calculation to the planning decision, improve the detection efficiency and reduce the energy consumption;

[0092] (3) By using the multi-probe body to detect the scenario and interact with the multi-agent reinforcement learning algorithm to train the decision-making and planning model, the accuracy of the collaborative detection decision-making and planning of multiple detectors can be improved, and the distance between sim2real can be reduced.

[0093] (4) Considering the problem of the sparsity of the grid probability in point cloud computing, a visual detection probability distribution function of the asteroid surface based on the weighted detection sensitivity function is proposed, which solves the problem of difficult solution of the grid visual detection probability caused by the sparse sampling point cloud of the detector within the field of view.

[0094] (5) Aiming at the problem of evaluating the decision-making and planning of the three-dimensional reconstruction detection of multiple spacecraft on the planet, a reward function based on the overall grid gain variance of the asteroid surface is proposed, which can ensure the execution efficiency of the multi-spacecraft detection decision-making and planning method trained by multi-agent reinforcement learning while ensuring the overall effect of the three-dimensional reconstruction. Description of the Drawings

[0095] Figure 1 It is a schematic diagram of the asteroid and the coordinate system of multiple detectors;

[0096] Figure 2 It is a schematic diagram of the asteroid detection environment of the multi-probe body;

[0097] Figure 3 It is a schematic diagram of the rasterization of the asteroid concept envelope;

[0098] Figure 4 It is a flow chart of the collaborative detection decision-making and planning of multiple detectors based on visual uncertainty feedback;

[0099] Figure 5 It is a schematic diagram of the ratio of the detection sensitivity within the detection field of view of the detector changing with the detection diameter. Detailed Embodiments

[0100] The following further describes the detailed embodiments of the present invention with reference to the drawings.

[0101] The present invention proposes a multi-probe collaborative asteroid detection embodied environment, uses the depth vision camera carried by the probe body to collect three-dimensional information of the local surface of the asteroid, then shares the perception information among multiple detectors, constructs a global grid uncertainty probability map of the local asteroid surface, and designs a reward function with final task effect feedback to achieve efficient and high-quality three-dimensional reconstruction of the asteroid by multiple detectors.

[0102] The prior art is mainly based on the detection technology of the Bennu asteroid by NASA. Its method requires prior knowledge of the approximate three-dimensional shape of the asteroid to perform polygon division, and then uses parameters such as the incident angle and exit angle of sunlight on each polygon as feedback for training to generate a detection decision-making and planning model. In view of the problem of insufficient generalization ability of this method for detecting unknown targets, the present invention proposes a method for calculating the probability of point cloud sampling of a depth camera in an embodied detection environment combined with a conceptual spherical grid envelope of an asteroid, which can effectively solve the problem of insufficient generalization ability of the current multi-detector asteroid detection method for detecting unknown asteroid targets.

[0103] The prior art does not have a quantitative evaluation and feedback of the overall task completion effect, so it cannot fully guarantee the completion degree of the task execution effect of the trained model. Considering this, the present invention proposes to use the standard deviation of the full-grid gain on the asteroid surface as a reward function, which can ensure the execution efficiency of the multi-spacecraft detection decision-making and planning method trained based on multi-agent reinforcement learning and the balance of the overall three-dimensional reconstruction effect.

[0104] As Figure 4 shown, the following is a specific introduction to the asteroid intelligent collaborative detection and planning method with visual uncertainty feedback proposed by the present invention:

[0105] 1) Conduct dynamic modeling of multi-detector asteroid detection and modeling of coordinate system conversion relationships

[0106] (1) Conduct dynamic modeling of multi-detector asteroid detection

[0107] As Figure 1 shown, a relative orbital coordinate system o-x is constructed with the centroid of the asteroid as the origin w y w z w and an asteroid body coordinate system o-x is constructed b y b z b , and multi-sub-star body coordinate systems o-x are constructed s1 y s1 z s1 、o-x s2 y s2 z s2 、o-x s3 y s3 z s3 …

[0108] Assume that the detector is a particle and satisfies the relative orbital dynamics equation as follows:

[0109]

[0110] where r represents the orbital radius; θ is the orbital rotation angle of the relative orbital coordinate system with respect to the space inertial coordinate system; x, y, and z respectively represent the three-axis coordinates of the satellite in the relative orbital coordinate system, and are the corresponding velocity and acceleration; f x , f y , f z is the acceleration of each axis of the detector in the relative orbital coordinate system, and its value is the sum of the perturbation acceleration and its own power; G x , G y , G z are the projections of the asteroid gravitational field on the three axes of the relative orbital coordinate system; μ is the gravitational constant of the two-body system.

[0111] The definition of the relative orbital coordinate system o-x w y w z w is: with the centroid of the asteroid as the origin, the direction from the centroid of the sun to the centroid of the asteroid is the x w axis, the direction of the asteroid's running speed is the y w axis, and the z w axis is the direction of the asteroid's angular momentum, which conforms to the right-hand rule with the x w axis and the y w axis;

[0112] The definition of the asteroid body coordinate system o-x b y b z b is: the x b axis is the direction of the maximum inertia axis of the asteroid, the z b axis is the direction of the minimum inertia axis of the asteroid, and the y b axis satisfies the right-hand rule with the x b axis and the y b axis;

[0113] The definitions of the multi-satellite body coordinate systems o-x s1 y s1 z s1 , o-x s2 y s2 z s2 , o-x s3 y s3 z s3 are the same as the definition of the asteroid body coordinate system.

[0114] The positive direction of the x-axis of the detector is the direction of the system's depth vision sensor. Here, we assume that at any moment, our detector is directed towards the centroid of the asteroid.

[0115] (2) Model the coordinate system transformation relationship

[0116] ① Conversion between the asteroid body coordinate system and the relative orbit coordinate system

[0117] Considering the rotation of the asteroid, a body coordinate system o-x b y b z b is constructed for the asteroid, and the relative orbit coordinate system o-x w y w z w are considered. Assuming that the z b axis of the body coordinate system coincides with the z w axis of the relative orbit coordinate system, the relationship between the body coordinate system and the relative orbit coordinate system is as follows:

[0118]

[0119] where η = η0 + ωt. η is the relationship between the x-axis of the body coordinate system and the x-axis of the relative orbit coordinate system, η0 is the initial value of this angle, ω is the rotational angular velocity of the asteroid body coordinate system relative to the orbit coordinate system, and t is the rotation time.

[0120] ② Conversion between the detector body coordinate system and the relative orbit coordinate system

[0121]

[0122] where α, β, and γ are the angles corresponding to the coincidence of the relative orbit coordinate system with the detector body coordinate system after rotating around its own three coordinate axes respectively. x si 、y yi 、z zi are the point coordinates in the detector i body coordinate system; x w 、y w 、z w are the point coordinates in the relative orbit coordinate system.

[0123] ③ Conversion between the detector and the asteroid body coordinate system

[0124]

[0125] where first convert the body coordinate system to the relative orbit coordinate system, and then convert it to the detector body coordinate.

[0126] 2) Construct a multi-detector collaborative in-situ detection scenario for asteroids

[0127] As Figure 2As shown in the figure, a physical engine such as GAZEBO is used as a platform for constructing an embodied detection scenario. The body coordinate system of the asteroid and the detection device body model is calibrated and implanted into the physical engine. At the same time, the conversion relationship between coordinate systems is set. On this basis, the body model of the depth camera is implanted, the camera coordinate system is calibrated, and the conversion between the camera coordinate system and the detector body coordinate system is completed.

[0128] 3) Construct the asteroid concept envelope model and grid division

[0129] As Figure 3 shown in the figure, regardless of the shape of the asteroid, it can be constructed as being enveloped by a spherical envelope with a certain diameter. A concept envelope of the asteroid is preset inside each spacecraft, and its surface is evenly divided into polygon grids.

[0130] The asteroid is enveloped in a spherical envelope. The radius of this envelope is the maximum radius of the asteroid, which is called the asteroid concept envelope, that is, the asteroid concept envelope model.

[0131] Envelope grid division: First, design that the asteroid concept envelope is composed of twenty spherical equilateral triangles. Then each spherical equilateral triangle is divided into 4 small spherical equilateral triangles, and based on this step, the spherical envelope is continuously divided. The granularity of the equilateral triangle primitive that makes up the asteroid concept envelope depends on the field of view radius determined by the flight altitude and field of view angle of the detector. The diameter of the envelope triangle primitive needs to be less than the detector field of view diameter. All the divided grids are the set U.

[0132] The environmental certainty status of each divided grid during the spacecraft detection process is represented by three states: occupied, unknown, and idle.

[0133] 4) Define detection visual uncertainty, environmental uncertainty, and information gain

[0134] (1) Define the visual uncertainty within the detection field of view

[0135] In the embodied detection environment, the spacecraft uses a depth camera to detect and sample, and obtains three-dimensional feature points on the surface of the asteroid within the field of view, that is, point cloud. The state probability of the positions marked by the point cloud is estimated. The proportion of the number of point cloud feature points falling into each grid of the envelope within the field of view to the total number of point clouds in this sampling is used as the visual uncertainty probability P of the grid (m,n). υ .

[0136] Considering that the point cloud acquisition by the depth camera is prone to the problem of sparse point cloud, the sensitivity distribution of the depth camera detection is modeled and weighted and summed with the probability initially calculated from the point cloud to obtain a more reasonable probability distribution of the visual uncertainty of the asteroid surface grid detection.

[0137]

[0138] Among them, M represents the sensitivity distribution function within the detection field of view of the detector;

[0139] L(x) = exp(x) / (1 + exp(x)); D is the boundary of the field of view. When approaching the upper boundary, that is, the outermost edge of the field of view of D, the sensitivity is 0; is the tolerance, reflecting the most sensitive range within the field of view; ζ1 and ζ2 are coefficients. For example Figure 5 is a schematic diagram of the change ratio of the detection sensitivity within the detection field of view of the detector with the detection diameter.

[0140] (2) Define the uncertainty of the asteroid surface environment

[0141] Assign the probability of the asteroid environment uncertainty to each grid (m,n) ∈ U in the grid set U of the asteroid concept envelope. During the detection process of the detector, update the environmental uncertainty probability of each grid using the visual uncertainty within the detection field of view of the detector. When the environmental uncertainty of the entire envelope grid set U is reduced to the lowest, the purpose of this detection is achieved. Represent the environmental uncertainty status of each grid on the asteroid surface during the spacecraft detection process in three states: occupied, unknown, and idle.

[0142] In the initial state, the asteroid surface environment is completely unknown, and the environmental uncertainty probability P occ (m,n) of all grids are set to 0.5, (m,n) ∈ U; when the probability reaches the set upper boundary P occ_threshold , the grid is considered occupied and no longer plan to reach this point repeatedly; when P occ (m,n) is lower than the lower boundary P free_threshold , it is idle, that is, try to plan a route to detect here as much as possible; for the area not detected by the detector, mark it as unknown, and P occ (m,n) remains 0.5.

[0143] The environmental uncertainty P mn (t) of a certain grid (m,n) on the asteroid surface is updated using the Bayesian probability update mechanism:

[0144] ① When the detector does not detect the grid (m,n)

[0145] P occ ((m,n),t) = P occ ((m,n),t - 1)

[0146] ② When the detector detects the grid (m,n)

[0147]

[0148] Among them, P drespectively represent the detection probability of the detector, i.e., visual uncertainty; P f represents the false alarm rate of faults; t represents the number of detection steps of the detector.

[0149] (3) Define the visual detection gain within the detection field of view

[0150] Suppose the detector detects the grid (m,n) on the asteroid surface. Due to the detector being at a certain height, all adjacent areas within the depth visual camera's detection field of view except the grid (m,n) will be observed, and the environmental grids outside the detection field of view will not be detected.

[0151] The environmental detection gain of the grid (m,n) is:

[0152] E(m,n) = -P occ (m,n)logP occ (m,n) - (1 - P occ (m,n))log(1 - P occ (m,n))

[0153] The visual detection gain of adjacent grids within its detection field of view is also calculated according to this formula. For areas that have not been detected at all during the entire mission process, it can be judged from this formula that the detection information gain is the largest, and thus the spacecraft can be guided to detect in areas that have not been detected.

[0154] 5) Multi-detector intelligent collaborative detection decision-making and planning based on visual uncertainty feedback

[0155] The detection process is a process of continuous entropy reduction. The detector needs to plan a path and act according to this path, and the entropy energy on the asteroid surface continuously decreases.

[0156] (1) Information sharing

[0157] To promote multi-detector collaboration, a globally consistent environmental model is built into each detector, and exploration planning is carried out based on this. During the detection process, each detector is initialized with the visual detection gain information of all grids on the asteroid surface, that is, the environmental detection gain map of the asteroid surface.

[0158] Through coordinate transformation of multiple detectors, all detector coordinates are converted into the asteroid body coordinate system. In this way, each detector can continuously maintain and update an environmental detection gain map of the asteroid surface with the asteroid body coordinate system as a reference. Each detector shares position / velocity data in real time and updates the environmental detection gain map of the asteroid surface after receiving the latest data. The advantage of this form is that unless communication fails, each detector has a complete local global uncertainty information map, can collect global information from the start to the end of the mission, and provides sufficient global information guarantee for multi-detector collaboration.

[0159] (2) Multi - detector Cooperative Detection Decision Planning Based on Multi - agent Reinforcement Learning

[0160] In order to reduce the uncertainty of the map while minimizing the cost, the problem can be summarized as finding a path that can observe more information about the environment at a relatively low cost.

[0161] Design an interaction training between the multi - agent reinforcement learning algorithm and the multi - detection equipment in the asteroid detection environment to generate a multi - spacecraft asteroid detection intelligent decision - making planning model. The action space, state space, and reward function are designed as follows.

[0162] ① Action Space

[0163] Design the action space as the three - axis accelerations [a x , a y , a z of each detector, where a x , a y , a z are all less than the maximum acceleration a max ;

[0164] ② State Space

[0165] [p i , v i , Δp ij , Δv ij ,…], i = 1, 2, …, j ≠ i, where p i is the position of detector i, v i is the velocity of detector i, Δp ij is the position difference between detector i and detector j, and Δv ij is the velocity difference between detector i and detector j;

[0166] ③ Reward Function

[0167] A. Collision Reward

[0168] In the embodied detection environment, if a detector collides with an asteroid or detectors collide with each other during the training process, the embodied detection environment gives negative feedback R1 and ends the current training;

[0169]

[0170] Among them, R1 represents the collision reward between the detector and the asteroid. When the detector is greater than the radius d w of the asteroid and less than a certain flight range d w +Δ s , then a positive reward μ is given; Δ sis a certain flight altitude range of the detector from the surface of the asteroid, p o represents the origin position in the asteroid body coordinate system, p i represents the position of the detector in the asteroid body coordinate system.

[0171] When the detector is smaller than the asteroid radius d w , a negative reward -κ is given for the collision between the detector and the asteroid;

[0172] When the flight range of the detector exceeds d w +Δ s , a negative guidance reward -τ·|p i -p o | is given to pull it back to the set detection range; τ is a coefficient.

[0173]

[0174] Among them, R2 is the collision reward between detectors, and k and σ are both positive coefficients. When detectors i and j are greater than each other's minimum safety threshold d th and smaller than the safety margin Δ ij , a positive reward is given to maintain this distance as much as possible; when detectors i and j are smaller than each other's minimum safety threshold d th , a negative reward is given, and the closer the approaching distance, the more severe the negative reward; p i and p j represent the positions of detectors i and j in the asteroid body coordinate system.

[0175] B. Guidance Reward

[0176] The detector tries to detect the direction with the largest information gain within the field of view as the guidance mechanism;

[0177] R3 = λ·|p i -Z max_i |

[0178] Among them, λ is a coefficient, p i represents the position corresponding to the i-th detector, and Z max_i represents the position with the largest information gain within the field of view detected by the position where the i-th detector is located; R3 is the guidance reward for the detector to move towards the direction with the largest information gain within the field of view.

[0179] C. 3D Reconstruction Effect Reward

[0180] During the detection process, the uniformity of all grid sampling point clouds on the surface of the asteroid can most significantly indicate the asteroid detection and reconstruction effect. The uniformity and quantity of the depth camera sampling point clouds are represented by the standard deviation of the full-grid gain. The specific reward function is set as follows:

[0181]

[0182] Among them, num is the number of grid divisions on the asteroid surface; S is the standard deviation; is a coefficient, and R4 is the reward for the 3D reconstruction effect. That is, the higher the standard deviation, the more uneven the overall point cloud coverage acquisition on the asteroid surface, and the lower the reward; is the sum of the grid gains divided by the number of grids.

[0183] The total reward function is as follows:

[0184] R all = ρ1·R1 + ρ2·R2 + ρ3·R3 + ρ4·R4

[0185] Among them, ρ1, ρ2, ρ3, and ρ4 are coefficients, and R1, R2, R3, and R4 respectively correspond to the above reward functions;

[0186] ④ Training and testing

[0187] Initialize the states of multiple detectors and network parameters. The multiple detectors update their states by interacting with the task environment and update the network parameters according to the state gradient descent of multi-agent reinforcement learning. When the reward function of the multiple detectors stabilizes within a certain range and no longer rises, stop training and save the model.

[0188] Test the trained decision network: Instantiate the multi-agent reinforcement learning network and import the trained decision model, initialize the detector and environment states. The detectors obtain the environmental observation information through their respective local perceptions as the input of the decision network, and the output is their respective accelerations. Update the detection plan through continuous iteration. Finally, end the test with the maximum environmental information gain no longer rising as the algorithm evaluation criterion.

[0189] The parts not detailed in the present invention belong to the common general knowledge of those skilled in the art.

Claims

1. An asteroid intelligent collaborative detection planning method with visual uncertainty feedback, characterized in that Including: Conduct dynamic modeling of multi-detector asteroid exploration and establish the coordinate system transformation relationship model; Construct a multi-detector collaborative asteroid embodied exploration scenario; Construct an asteroid conceptual envelope model and grid division; Define the detection visual uncertainty, environmental uncertainty, and information gain; Conduct intelligent collaborative detection decision-making planning for multi-detectors based on visual uncertainty feedback.

2. The method for asteroid intelligent collaborative detection planning based on visual uncertainty feedback according to claim 1, wherein: The conduct of dynamic modeling of multi-detector asteroid exploration is specifically as follows: Construct a relative orbital coordinate system o-x with the centroid of the asteroid as the origin w y w z w and construct the body coordinate system o-x of the asteroid b y b z b , construct the body coordinate systems of multiple sub-stars o-x s1 y s1 z s1 、o-x s2 y s2 z s2 、o-x s3 y s3 z s3 、…; Assume the detector as a particle and satisfy the relative orbit dynamics equation as follows: where r represents the orbital radius; θ is the orbital rotation angle of the relative orbital coordinate system with respect to the space inertial coordinate system; x, y, and z respectively represent the three-axis coordinates of the satellite in the relative orbital coordinate system. and are the corresponding velocity and acceleration; f x , f y , f z is the acceleration of each axis of the detector in the relative orbital coordinate system, and its value is the sum of the perturbation acceleration and its own power; G x , G y , G z are the projections of the gravitational field of the asteroid on the three axes of the relative orbital coordinate system; μ is the gravitational constant of the two-body system.

3. According to the method for asteroid intelligent collaborative detection planning based on visual uncertainty feedback described in claim 2, wherein: Relative orbital coordinate system o-x w y w z w is defined as follows: with the centroid of the asteroid as the origin, the direction from the centroid of the sun to the centroid of the asteroid is the x w axis, the direction of the running speed of the asteroid is the y w axis, the z w axis is the direction of the angular momentum of the asteroid, which conforms to the right-hand rule with the x w axis and the y w axis; The body coordinate system o-x of the asteroid b y b z b is defined as follows: The x b axis is in the direction of the maximum inertia axis of the asteroid, the z b axis is in the direction of the minimum inertia axis of the asteroid, and the y b axis, together with the x b axis and the y b axis, satisfies the right-hand rule; Multi-asteroid body coordinate system o-x s1 y s1 z s1 、o-x s2 y s2 z s2 、o-x s3 y s3 z s3 The definition of is the same as that of the asteroid body coordinate system.

4. A method for asteroid intelligent collaborative detection planning with visual uncertainty feedback according to claim 1, characterized in that: The establishment of the coordinate system transformation relationship model specifically includes: ① Transformation between the asteroid body coordinate system and the relative orbit coordinate system Considering the rotation of the asteroid, a body coordinate system o-x of the asteroid is constructed b y b z b The transformation relationship with the relative orbital coordinate system o-x w y w z w is considered. Assuming that the z-axis of the asteroid body coordinate system b coincides with the z-axis of the relative orbital coordinate system w the relationship between the asteroid body coordinate system and the relative orbital coordinate system is as follows: Where η = η0 + ωt, η is the angle between the x-axis of the asteroid body coordinate system and the x-axis of the relative orbit coordinate system, η0 is the initial value of this angle, ω is the rotational angular velocity of the asteroid body coordinate system relative to the orbit coordinate system, and t is the rotation time; ② Transformation between the detector body coordinate system and the relative orbit coordinate system where α, β, and γ are the angles corresponding to the coincidence of the relative orbital coordinate system with the detector body coordinate system after rotating around the three axes of its own coordinate system; x si , y yi , z zi are the point coordinates in the detector i body coordinate system; x w , y w , z w are the point coordinates in the relative orbital coordinate system; ③ Transformation between the detector body coordinate system and the asteroid body coordinate system Among them, first convert the asteroid body coordinate system to the relative orbital coordinate system, and then convert it to the detector body coordinate system.

5. A method for asteroid intelligent cooperative detection planning with visual uncertainty feedback according to claim 1, characterized in that: The construction of the multi-detector collaborative asteroid embodied exploration scenario is specifically as follows: Construct the GAZEBO physics engine as the platform for constructing the embodied exploration scenario, calibrate the body coordinate systems of the asteroid and the exploration instrument body models and implant them into this physics engine, and at the same time set the transformation relationship between the coordinate systems; on this basis, implant the embodied model of the depth camera, calibrate the camera coordinate system and complete the transformation between the camera coordinate system and the detector body coordinate system.

6. The intelligent collaborative detection planning method for asteroids with visual uncertainty feedback according to claim 1, wherein: The construction of the asteroid conceptual envelope model and grid division is specifically as follows: Envelop the asteroid in a spherical envelope body, the radius of which is the maximum radius of the asteroid, called the asteroid conceptual envelope body, that is, the asteroid conceptual envelope model; Grid division of the envelope body: First, design that the asteroid conceptual envelope body is composed of twenty spherical equilateral triangles, and then each spherical equilateral triangle is divided into 4 small spherical equilateral triangles, and the spherical envelope body is continuously divided based on this step; The size of the elementary particle of the equilateral triangle that constitutes the asteroid conceptual envelope body depends on the field of view radius determined by the detector flight height and the field of view angle, and the diameter of the triangular elementary unit of the envelope body needs to be less than the detector field of view diameter; all the divided grids are the set U.

7. An asteroid intelligent collaborative detection planning method with visual uncertainty feedback according to claim 6, characterized in that: The definition of the detection visual uncertainty, environmental uncertainty, and information gain specifically includes: (4.1) Define the visual uncertainty within the detection field of view In the embodied detection environment, the spacecraft uses a depth camera to detect and sample, and obtains three-dimensional feature points on the surface of the asteroid within the field of view, that is, point clouds. The state probability of the positions marked by the point clouds is estimated. The ratio of the number of point cloud feature points falling in each grid of the in-field envelope to the total number of point clouds in this sampling is used as the visual uncertainty probability P of the grid (m,n). υ ; Model the sensitivity distribution of the depth camera detection and perform weighted summation with the probability obtained from the preliminary calculation of the point cloud to obtain a more reasonable probability distribution of the detection visual uncertainty of the asteroid surface grid Where M represents the sensitivity distribution function within the detector detection field of view range; L(x) = exp(x) / (1 + exp(x)); D is the field of view boundary. When approaching the upper boundary, that is, the outermost edge of the field of view of D, the sensitivity is 0; is the tolerance, reflecting the most sensitive range within the field of view; ζ1 and ζ2 are coefficients; (4.2) Define the environmental uncertainty of the asteroid surface Assign the probability of asteroid environmental uncertainty to each grid (m,n) ∈ U in the asteroid concept envelope grid set U. During the detection process of the detector, update the environmental uncertainty probability of each grid using the visual uncertainty within the detection field of view of the detector. When the environmental uncertainty of the entire envelope grid set U drops to the lowest, the purpose of this detection is achieved. Represent the environmental uncertainty status of each grid on the asteroid surface during the spacecraft detection process in three states: occupied, unknown, and idle. In the initial state, the surface environment of the asteroid is completely unknown, and the uncertainty probability P of all grid environments occ (m,n) is set to 0.5 for all (m,n) ∈ U; when the probability reaches the set upper bound P occ_threshold , the grid is considered occupied and the route planning to this point is not repeated; when P occ (m,n) is lower than the lower bound P free_threshold , it is free, that is, the route should be planned as much as possible to explore this area; for the areas not detected by the detector, they are marked as unknown and P occ (m,n) remains 0.5; Environmental uncertainty P of the asteroid surface grid (m, n) mn (t) The update adopts a Bayesian probability update mechanism: ① When the detector does not detect the grid (m,n) P occ ((m,n),t) = P occ ((m,n),t - 1) ② When the detector detects the grid (m,n) Among them, P d respectively represents the detection probability of the detector, i.e., the visual uncertainty; P f represents the false alarm rate of the fault; t represents the detection steps of the detector; (4.3) Define the visual detection gain within the detection field of view Suppose the detector detects the grid (m,n) on the asteroid surface. Due to the detector being at a certain height, all adjacent areas within the depth visual camera's detection field of view except the grid (m,n) will be observed, and the environmental grids outside the detection field of view will not be detected. The environmental detection gain of the grid (m,n) is: E(m,n) = -P occ (m,n) log P occ (m,n) - (1 - P occ (m,n)) log(1 - P occ (m,n)) The visual detection gain of adjacent grids within its detection field of view is also calculated according to this formula. For areas that have not been detected at all during the entire mission process, the formula determines that the detection information gain is the largest, and thus the spacecraft is guided to detect the undetectable areas.

8. A method for asteroid intelligent collaborative exploration planning with visual uncertainty feedback according to claim 7, characterized in that: The multi-detector intelligent collaborative detection decision-making and planning based on visual uncertainty feedback is specifically as follows: (5.1) Conduct information sharing During the detection process, each detector is initialized with the visual detection gain information of all grids on the asteroid surface, that is, the environmental detection gain map of the asteroid surface. During the detection process, each detector will share information with each other and continuously update this map according to the visual uncertainty, environmental uncertainty within the detection field of view, and the visual detection gain calculation formula, so as to dynamically guide the detector to grids with high detection gain. (5.2) Conduct multi-detector collaborative asteroid detection decision-making and planning based on multi-agent reinforcement learning, that is, design an interactive training between the multi-agent reinforcement learning algorithm and the multi-detection device's own asteroid detection environment to generate a multi-spacecraft asteroid detection intelligent decision-making and planning model. The action space, state space, and reward function are designed as follows; ① Action space The designed action space is the three-axis accelerations [a x , a y , a z of each detector, where a x , a y , a z are all less than the maximum acceleration a max ; ② State space [p i , v i , Δp ij , Δv ij , …], i = 1, 2, …, j ≠ i, where p i is the position of detector i, v i is the velocity of detector i, Δp ij is the position difference between detector i and detector j, Δv ij is the velocity difference between detector i and detector j; ③ Reward function A. Collision reward In the embodied detection environment, if the detector collides with the asteroid or the detectors collide with each other during the training process, the embodied detection environment gives a negative feedback R1 and ends this training; Among them, R1 represents the collision reward between the detector and the asteroid. When the detector is larger than the radius d of the asteroid w and smaller than a certain flight range d w +Δ s , then a positive reward μ is given; Δ s is a certain flight height range of the detector from the surface of the asteroid, and p o represents the origin position in the asteroid's body coordinate system, and p i represents the position of the detector in the asteroid's body coordinate system; When the detector is smaller than the radius d of the asteroid w , a negative reward -κ is given for the collision between the detector and the asteroid; When the flight range of the detector exceeds d w +Δ s , a negative guidance reward of -τ·|p i -p o | will be given to pull it back to the set detection range; τ is a coefficient; Among them, R2 is the collision reward between detectors, and k and σ are both positive coefficients. When the detectors i and j are greater than each other's lowest safety threshold d th and less than the safety margin Δ ij , a positive reward is given to maintain this distance as much as possible; when the detectors i and j are less than each other's lowest safety threshold d th , a negative reward is given, and the closer the approaching distance is, the more severe the negative reward is; p i and p j represent the positions of the detectors i and j in the asteroid body coordinate system; B. Guidance reward The detector tries to detect the direction with the largest information gain within the detection field of view as the guidance mechanism; R3 = λ·|p i -Z max_i | where λ is a coefficient, p i represents the position corresponding to the i-th detector, and Z max_i represents the position with the maximum information gain within the field of view detected at the position where the i-th detector is located; R3 is the guiding reward for the detector to move in the direction with the maximum information gain within the detected field of view; C. 3D reconstruction effect reward The uniformity of the sampling point clouds of all grids on the asteroid surface during the detection process can most indicate the asteroid detection reconstruction effect. The uniformity and quantity of the depth camera sampling point clouds are represented by the standard deviation of the full-grid gain. The specific reward function is set as follows: where num is the number of grid divisions on the asteroid surface; S is the standard deviation; is a coefficient, and R4 is the reward for the 3D reconstruction effect. That is, the higher the standard deviation, the more uneven the overall point cloud coverage collection on the asteroid surface, and the lower the reward; is the sum of the grid gains divided by the number of grids; The total reward function is as follows: R all = ρ1·R1 + ρ2·R2 + ρ3·R3 + ρ4·R4 Among them, ρ1ρ2ρ3ρ4 are coefficients, and R1, R2, R3, R4 correspond to the above reward functions respectively; ④ Training and testing Initialize the states of multiple detectors and the parameters of the decision network. The multiple detectors update their states by interacting with the task environment, and update the parameters of the decision network according to the state gradient descent of multi-agent reinforcement learning. When the reward function of the multiple detectors stabilizes within a certain range and no longer increases, stop the training and save the model as the collaborative asteroid detection decision-making and planning model for multiple detectors.

9. The method for asteroid intelligent collaborative detection planning with visual uncertainty feedback according to claim 8, wherein: Test the trained collaborative asteroid detection decision-making and planning model for multiple detectors: instantiate the multi-agent reinforcement learning network and import the trained decision-making and planning model, initialize the states of multiple detectors and the environment. The multiple detectors obtain the environmental observation information through their respective local perceptions as the input of the decision network, output their respective accelerations, and update the detection plan through continuous iteration. Finally, end the test with the maximum environmental information gain no longer increasing as the algorithm evaluation criterion.

Images