Unmanned aerial vehicle interpretable autonomous maneuver penetration method

Abstract

The invention discloses an interpretable autonomous maneuver penetration method for an unmanned aerial vehicle, and the method comprises the following steps: building a maneuver element, and forming a three-dimensional discrete motion space; through continuous action learning automaton training, the probabilities of different maneuvering primitives are updated, a lookup table is generated, and the lookup table is carried on the unmanned aerial vehicle; and the unmanned aerial vehicle queries the lookup table according to the state of the unmanned aerial vehicle to obtain a corresponding maneuvering element, and the unmanned aerial vehicle performs maneuvering based on the obtained maneuvering element to realize penetration. The autonomous maneuver penetration method capable of being explained by the unmanned aerial vehicle disclosed by the invention has remarkable advantages in the aspects of maneuver adaptability, robustness and airborne energy efficiency.

Description

[0001] The application relates to an interpretable autonomous maneuvering penetration method for a UAV, and belongs to the technical field of aircraft control. BACKGROUND

[0002] A UAV attempting to penetrate a layered air defense network must cope with multiple interceptors and complex no-fly zones, and needs agile and optimal evasive maneuvers under strict aircraft constraints.

[0003] Traditional penetration methods, including performing a wiggle or oscillation maneuver, a sinusoidal evasive trajectory, terminal maneuver penetration, etc., cannot adapt to unpredictable, highly nonlinear engagement dynamics. To solve the above problems, methods for path planning using artificial intelligence are widely studied.

[0004] Methods for path planning using artificial intelligence to achieve penetration, such as memory-augmented deep Q networks and distributed policy optimization methods, these artificial agents learn through trial and error, perform sudden dives, sharp turns and large-overload turns, and achieve good penetration performance, however, this powerful capability is accompanied by a trade-off: end-to-end neural policies often behave like an "uninterpretable black box" and provide almost no formal safety guarantees.

[0005] Therefore, it is necessary to obtain the interpretability of the learning decision-making process of the agent, so that human operators can understand and trust the behavior of autonomous agents. SUMMARY

[0006] In order to overcome the above problems, in-depth research is conducted, and an interpretable autonomous maneuvering penetration method for a UAV is proposed, comprising the following steps:

[0007] S1, establishing a maneuvering primitive and forming a three-dimensional discrete action space;

[0008] S2, updating the probabilities of different maneuvering primitives through continuous action learning automaton training, generating a lookup table, and loading the lookup table on the UAV;

[0009] S3, the UAV queries the lookup table according to its state to obtain the corresponding maneuvering primitive, and the UAV performs maneuvering based on the obtained maneuvering primitive to achieve penetration.

[0010] The maneuvering primitive includes one or more of straight-line acceleration, U-turn, S-turn, square wave, barrel roll and inclined plane;

[0011] The straight-line acceleration refers to only increasing the longitudinal velocity;

[0012] The U-turn refers to a roll action achieved by applying significant vertical and lateral acceleration;

[0013] The S-turn refers to using a sinusoidal form of lateral acceleration to produce an unpredictable flight trajectory;

[0014] The square wave refers to achieving sudden defense by periodically performing sudden direction changes;

[0015] The barrel roll refers to combining rolling motion with sustained lateral acceleration;

[0016] The inclined surface refers to achieving moderate-intensity continuous direction changes along an inclined trajectory.

[0017] In a preferred embodiment, the actions of the maneuver primitives are combined to form a three-dimensional discrete action space, in which the action primitive parameters are represented in the format of a three-dimensional action triplet, and the three-dimensional discrete action space is represented as:

[0018]

[0019] Wherein, a represents the three-dimensional discrete action space, M type represents a set of different maneuver primitive sequence numbers, n set represents a set of maximum overload amplitudes applied to the maneuver primitives, which are used to amplify or reduce the overload in the maneuver primitives, r set represents a set of distance threshold values for triggering the maneuver primitives.

[0020] In a preferred embodiment, the three-dimensional discrete action space is discretized into a plurality of tensor grid structures, which can be represented as:

[0021]

[0022] Wherein, G represents the tensor node, i represents the maneuver primitive sequence number, N m represents the total number of maneuver primitives, M i represents the i-th maneuver primitive, j represents the maximum overload amplitude sequence number, N n represents the total number of set maximum overload amplitudes, n j represents the j-th maximum overload amplitude, k represents the distance threshold value sequence number for triggering the maneuver primitives, N r represents the total number of set distance threshold values, r k represents the k-th distance threshold value.

[0023] In a preferred embodiment, in the training, the probabilities of different maneuver primitives are updated based on simulation, including the following sub-steps:

[0024] S21, at the beginning of each round of training, a scene is randomly generated, and the continuous action learning automaton selects a tensor node from the current probability density tensor by the method of sampling from the inverse cumulative distribution function, and simulates based on the maneuver primitive, maximum overload amplitude and distance threshold in the tensor node to obtain the survival probability of the UAV through simulation;

[0025] S22, the probability density is updated according to the survival probability obtained in this round of training;

[0026] S23, repeat the above process to perform multiple iterations of training and update all probability density tensors;

[0027] S24, normalize the entire probability density tensor to form a lookup table.

[0028] In a preferred embodiment, in S21, the selected tensor node is simulated multiple times, and the average survival probability obtained from multiple simulations is taken as the survival probability of the UAV in this round of training.

[0029] In a preferred embodiment, at the initial moment, the probability density tensor is set to be uniformly distributed.

[0030] In a preferred embodiment, S22, the probability density is updated according to the survival probability obtained in this round of training, which is represented as:

[0031]

[0032] K j,k =K n (j)K r (k)

[0033]

[0034] wherein, denotes the updated probability density tensor, η S denotes the learning rate of the reward, R t denotes the survival probability obtained in this round of training, η F denotes the learning rate of the penalty, P t (M t , j, k) denotes the probability density of maneuver primitive M t at the jth maximum overload amplitude and the kth distance threshold, K j,k , K r (k), K n (j) is a process function, r t denotes the distance threshold at time t, n t denotes the maximum overload amplitude at time t, σ r , σ nwidth parameter of the Gaussian kernel function.

[0035] In a preferred embodiment, in S23, the width parameter of the Gaussian kernel function r , σ n is gradually reduced.

[0036] In a preferred embodiment, in S3, the UAV further acquires the distance threshold and the maximum magnitude of overload triggering the maneuver primitive according to the query lookup table.

[0037] The present application has the beneficial effects including:

[0038] (1) The maneuver primitives are mapped onto a three-dimensional action grid, resulting in a strategy that is both intuitive and visually transparent;

[0039] (2) A large-scale randomized engagement scenario is used to generate a Monte Carlo replay: for each sampled action, its probability density is locally reinforced or suppressed, and the iterative convergence to a sharp peak around the optimal parameter combination is searched for, minimizing the dependence on the deep network architecture and hyperparameters, ensuring controlled convergence, and producing results that are easy to audit;

[0040] (3) The computational burden is reduced, and the real-time response capability is improved, enabling sub-millisecond delays on embedded processors;

[0041] (4) Significant advantages in maneuver adaptability, robustness, and onboard energy efficiency. BRIEF DESCRIPTION OF DRAWINGS

[0042] Figure 1 Fig. 1 shows a flowchart of a UAV interpretable autonomous maneuver penetration method according to a preferred embodiment of the present application;

[0043] Figure 2 Fig. 3 shows the comparison results of Example 1 and Comparative Example 1. DETAILED DESCRIPTION

[0044] The present application will be further described in detail by the accompanying drawings and examples. Through these descriptions, the features and advantages of the present application will become clearer and more explicit.

[0045] The term "exemplary" is used herein to mean "serving as an example, instance, or illustration." Any implementation described herein as "exemplary" is not necessarily to be construed as preferred or advantageous over other implementations. Unless specifically indicated otherwise, the drawings shown in the Figures are not necessarily drawn to scale.

[0046] According to the UAV interpretable autonomous maneuver penetration method provided by the present application, as shown in Fig. 1, the method comprises the following steps: Figure 1

[0047] ​S1, establish a maneuvering primitive and form a three-dimensional discrete action space;

[0048] S2, update the probabilities of different maneuvering primitives through continuous action learning automaton training to generate a lookup table;

[0049] S3, the UAV queries the lookup table according to its state to obtain the corresponding maneuvering primitive and trigger distance threshold, and the UAV performs a maneuver based thereon to achieve penetration.

[0050] In S1, the action space is constrained by using a predefined maneuvering primitive library meaningful to human pilots, so that the subsequent control output is no longer a direct control quantity, but a maneuvering primitive (e.g., left sharp turn, barrel roll, climb turn, etc.), making the strategy easier to interpret. Research shows that using a limited set of known maneuvering primitives during agent learning can produce human-understandable and verifiable combat tactics.

[0051] In a preferred embodiment, the maneuvering primitives include one or more of straight acceleration, U-turn, S-turn, square wave, barrel roll, and inclined plane.

[0052] The straight acceleration refers to only increasing the longitudinal velocity, represented as:

[0053]

[0054] wherein, represents the x-direction overload command, represents the y-direction overload command, represents the z-direction overload command, and g represents the gravitational acceleration.

[0055] The U-turn refers to achieving a roll action by applying significant vertical and lateral acceleration, represented as:

[0056]

[0057] wherein, θ represents the flight path inclination, and γ C represents the roll angle command for maneuvering in the horizontal and vertical planes.

[0058] The S-turn refers to using a sinusoidal form of lateral acceleration to produce an unpredictable flight trajectory, represented as:

[0059]

[0060] wherein, t represents the time variable, and γ represents the roll angle for maneuvering in the horizontal and vertical planes.

[0061] The square wave refers to achieving penetration by periodically performing sudden changes in direction, represented as:

[0062]

[0063] Rolling, refers to the combination of rolling motion and continuous lateral acceleration, expressed as:

[0064]

[0065] Inclined surface, refers to the realization of moderate intensity continuous direction change along the inclined trajectory, expressed as:

[0066]

[0067] These maneuver primitives enable the UAV to perform precise and efficient evasion actions without continuously recalculating detailed control inputs in real time, significantly reducing computational complexity and enhancing maneuver performance, improving real-time response capabilities.

[0068] Preferably, the actions of the maneuver primitives are combined to form a three-dimensional discrete action space, in which the action primitive parameters are represented in a three-dimensional action triplet format, and the three-dimensional discrete action space is represented as:

[0069]

[0070] Where a represents the three-dimensional discrete action space, M type represents a set of different maneuver primitive sequence numbers, n set represents a set of maximum overload amplitudes applied to the maneuver primitives, which are used to amplify or reduce the overload in the maneuver primitives, r set represents a set of distance thresholds for triggering the maneuver primitives, preferably, M type ∈{1,...,6}, n set ∈[9,12]g, r set ∈[1,8]km.

[0071] Once the UAV makes a decision based on the engagement situation and the relative distance reaches the trigger distance, the UAV performs the maneuver corresponding to the selected maneuver primitive at the specified overload amplitude. In this way, the UAV can determine the timing and manner of the maneuver according to the tactical situation.

[0072] According to the present application, the three-dimensional discrete action space is discretized into a plurality of tensor grid structures, which can be represented as:

[0073]

[0074] Where G represents the tensor node, i represents the maneuver primitive sequence number, N m represents the total number of maneuver primitives, M irepresents the i-th maneuvering element, j represents the maximum overload amplitude sequence number, N n represents the total number of set maximum overload amplitudes, n j represents the j-th maximum overload amplitude, k represents the distance threshold sequence number of triggering the maneuvering element, N r represents the total number of set distance thresholds, r k represents the k-th distance threshold.

[0075] In S2, the continuous action learning automaton (CARLA) is a framework developed by Howell et al. for adaptive control tasks, in which the classical stochastic learning automaton theory is extended to the continuous action domain by iteratively updating the probability distribution on the action space. In the present application, the specific structure of the continuous action learning automaton is not described in detail.

[0076] According to the present application, in S2, the continuous action learning automaton is used to update the probability density tensor under a fixed cell size.

[0077] Further, the probability density tensor under the fixed cell size can be expressed as:

[0078]

[0079] where P t (i,j,k) represents the probability density of the i-th maneuvering element at the j-th maximum overload amplitude and the k-th distance threshold at time t, and ΔrΔn represents the fixed cell size.

[0080] Preferably, at the initial time, the probability density tensor is set as a uniform distribution to reflect the lack of prior knowledge of the action space.

[0081] According to the present application, in the training, the probabilities of different maneuvering elements are updated based on simulation.

[0082] Specifically, the following sub-steps are included:

[0083] S21, at the beginning of each round of training, a scene is randomly generated, and the continuous action learning automaton selects a tensor node from the current probability density tensor by the method of inverse cumulative distribution function sampling, and simulates based on the maneuvering element, maximum overload amplitude and distance threshold in the tensor node to obtain the survival probability R t of the UAV through simulation, i.e. whether the penetration is successful, if the UAV survives, it means that the penetration is successful, when the UAV survives, the survival probability R t is 1, otherwise the survival probability R t of the UAV is 0.

[0084] According to the present application, in the randomly generated scene, the relative distance, line of sight angle and relative velocity of the UAV and the interceptor aircraft are randomly set.

[0085] Preferably, in S21, multiple simulations are performed using the selected tensor node, and the survival probability R t The mean value is taken as the survival probability of the UAV in this round of training.

[0086] Preferably, the Monte Carlo repeated experiment method is used for multiple simulations.

[0087] S22, the survival probability obtained in this round of training is used to update the probability density, denoted as:

[0088]

[0089] K j,k = K n (j) K r (k)

[0090]

[0091] wherein, denotes the updated probability density tensor, η S denotes the learning rate of the reward, R t denotes the survival probability obtained in this round of training, η F denotes the learning rate of the penalty, P t (M t , j, k) denotes the maneuver element M t at the jth maximum overload amplitude and the kth distance threshold, K j,k , K r (k), K n (j) is a process function, r t denotes the distance threshold at time t, n t denotes the maximum overload amplitude at time t, σ r , σ n is a width parameter of a Gaussian kernel function.

[0092] Further, when the updated probability density tensor has a negative value item, it is set to zero.

[0093] The above updating method makes the high-performance strategy corresponding region will be strengthened, and the low-performance region will be suppressed.

[0094] S23, repeat the above process for multiple iterations of training, and update all probability density tensors;

[0095] Preferably, in each update, the width parameter σ r , σ n of the Gaussian kernel function gradually decreases, and more preferably adopts a geometric decreasing method to gradually converge, denoted as:

[0096] σ r ←γσ r ,σ n ←γσ n ,0<γ<1.

[0097] where γ is a settable parameter.

[0098] Through the above setting, extensive exploration is realized in the early stage of training, and fine optimization of the optimal strategy is focused on in the later stage, so as to balance exploration and utilization.

[0099] S24, normalize the entire probability density tensor to form a lookup table.

[0100] The normalization is represented as:

[0101]

[0102] According to the present application, after multiple iterations of training, the probability density will form a clear peak on the optimal trigger condition and overload amplitude combination. These peaks are converted into a lookup table.

[0103] In S3, preferably, the unmanned aerial vehicle further acquires the distance threshold and the maximum overload amplitude of the triggered maneuver element according to the query lookup table.

[0104] In S3, the unmanned aerial vehicle keeps the sensor on during flight, and when detecting the presence of an interceptor aircraft, the unmanned aerial vehicle acquires the relative distance, line-of-sight angle, and relative speed through the sensor, acquires the maneuver element corresponding to the highest probability probability density tensor, the distance threshold and the maximum overload amplitude of the triggered maneuver element by querying the lookup table, and relies on these parameters to quickly enter the selected maneuver element at the right time.

[0105] Further, as the geometric situation evolves, the unmanned aerial vehicle will continuously monitor its own state; if the state changes, it can seamlessly perform the next lookup table query to ensure coherent response.

[0106] According to the method in the present application, the unmanned aerial vehicle does not perform any online learning during flight, only needs to perform a table lookup and simple maximum overload amplitude scaling, and does not require additional neural network hardware support, thereby controlling the decision delay within milliseconds, thereby realizing real-time control and high stability.

[0107] Embodiment

[0108] Embodiment 1

[0109] An experiment is performed using the interpretable autonomous maneuver penetration method of the unmanned aerial vehicle, including the following steps:

[0110] S1, establish a maneuver element and form a three-dimensional discrete action space;

[0111] S2, updating probabilities of different maneuver primitives by continuous action learning automaton training, generating a lookup table, and loading the lookup table on the UAV;

[0112] S3, the UAV querying the lookup table according to its state to obtain corresponding maneuver primitives, and the UAV maneuvering based on the obtained maneuver primitives to achieve penetration.

[0113] The maneuver primitives include straight acceleration, U-turn, S-turn, square wave, barrel roll, and inclined plane.

[0114] The maneuver primitives are combined to form a three-dimensional discrete action space, in which the action primitive parameters are represented in a three-dimensional action triplet format, and the three-dimensional discrete action space is represented as:

[0115]

[0116] The three-dimensional discrete action space is discretized into a plurality of tensor grid structures, which can be represented as:

[0117]

[0118] In the training, the probabilities of different maneuver primitives are updated based on simulation, including the following sub-steps:

[0119] S21, at the beginning of each round of training, a scene is randomly generated, and the continuous action learning automaton selects a tensor node from the current probability density tensor by the method of inverse cumulative distribution function sampling, simulates based on the maneuver primitives, maximum overload amplitude, and distance threshold in the tensor node, and obtains the survival probability of the UAV through simulation;

[0120] S22, updating the probability density according to the survival probability obtained in this round of training;

[0121] S23, repeating the above process for multiple iterations to update all probability density tensors;

[0122] S24, normalizing the entire probability density tensor to form a lookup table.

[0123] In S21, the selected tensor node is simulated multiple times, and the average survival probability obtained from multiple simulations is used as the survival probability of the UAV in this round of training.

[0124] At the initial moment, the probability density tensor is set to be uniformly distributed.

[0125] S22, updating the probability density according to the survival probability obtained in this round of training, which is represented as:

[0126]

[0127] K j,k = K n (j) K r (k)

[0128]

[0129] where η S is set to 0.6, η F is set to 0.3, P t (M t , j, k) represents the probability density of the jth maximum overload amplitude and the kth distance threshold at time t, K t , K j,k , K r (k), K n (j) are process functions, r t represents the distance threshold at time t, n t represents the maximum overload amplitude at time t, σ r , σ n are the width parameters of the Gaussian kernel function.

[0130] In S23, the width parameters σ r , σ n of the Gaussian kernel function are gradually reduced in each update, and are set as:

[0131] σ r ← γσ r , σ n ← γσ n , 0 < γ < 1.

[0132] γ is set to 0.995.

[0133] In S3, the UAV also obtains the distance threshold and the maximum overload amplitude of the triggered maneuver element according to the query of the lookup table.

[0134] In the simulation process, the fourth-order Runge-Kutta method is used for dynamic integration. To ensure that the scene is both realistic and widely covered, each round of training starts under new random initial conditions: the initial positions, velocities and attitude angles of the target and the interceptor are uniformly sampled within the preset ranges. The complete value range of all parameters is shown in Table 1.

[0135] Table 1

[0136]

[0137]

[0138] Once simulation ends under any one of the following three triggering conditions:

[0139] 1) Hit: relative distance R < 3m.

[0140] 2) Miss: the missile flies over the target and the distance exceeds 100m, and is not close again.

[0141] 3) Timeout: the simulation time reaches 60 seconds, and neither of the above two cases occurs.

[0142] The test is performed in 1000 brand-new combat scenarios, the parameters of which are consistent with those in the training.

[0143] Comparative Example 1

[0144] The test is performed in 1000 combat scenarios identical to those in Example 1, using a completely random parameter strategy.

[0145] The test results in Comparative Example 1 and Comparative Example 2 are compared, as shown in Table 1. Figure 2 As shown in Table 1, the penetration success rate of Example 1 is 88.8%, while the maneuvering strategy with random parameters is only 38.5%, with a difference of 50.3 percentage points between the two. With a sample size of 1000, this difference is statistically significant.

[0146] The results show that, although the rule-based maneuvering strategy in Comparative Example 1 can improve the survivability of the UAV to some extent, its success rate fluctuates greatly and is unstable in different situations; in contrast, the method in Example 1 improves the success rate by more than two times by selecting the optimal maneuvering method at the most appropriate time.

[0147] The above describes the present application in combination with preferred embodiments, but these embodiments are only exemplary and serve only to illustrate. On this basis, various substitutions and improvements can be made to the present application, and these all fall within the protection scope of the present application.

Claims

1. A method for autonomous maneuvering penetration of a UAV, characterized in that: The following steps are involved: S1. Establish maneuver primitives and form a three-dimensional discrete action space; S2, through continuous action learning automaton training, update the probabilities of different maneuver primitives, generate a lookup table, and carry the lookup table on the UAV; S3. The UAV queries the lookup table according to its status and obtains the corresponding maneuver primitive. The UAV performs maneuvers based on the obtained maneuver primitive to achieve penetration.

2. The interpretable autonomous maneuver penetration method of a UAV according to claim 1, characterized in that: The maneuver primitives include one or more of straight-line acceleration, U-turn, S-turn, square wave, barrel roll, and inclined plane; The linear acceleration mentioned herein refers to increasing only the longitudinal speed; The U-turn is a rolling action achieved by applying significant vertical and lateral accelerations; The S-turn refers to the use of sinusoidal lateral acceleration to produce an unpredictable flight trajectory; The square wave is to achieve penetration by periodically performing sudden changes in direction; The barrel roll is a combination of rolling motion and sustained lateral acceleration. The inclined surface refers to a continuous direction change of moderate intensity along an inclined trajectory.

3. The interpretable autonomous maneuver penetration method of a UAV according to claim 1, characterized in that: The actions of the maneuver primitives are combined to form a three-dimensional discrete action space. In the three-dimensional discrete action space, the action primitive parameters are expressed in the three-dimensional action triple format. The three-dimensional discrete action space is expressed as: Where a represents the three-dimensional discrete action space, M type Represents a set of different maneuver primitive numbers, n set represents the set of maximum overload amplitudes applied to the maneuver primitives, the maximum overload amplitudes being used to amplify or reduce the overload in the maneuver primitives, r set Represents a set of distance thresholds that trigger maneuver primitives.

4. The interpretable autonomous maneuver penetration method of a UAV according to claim 3, characterized in that: The three-dimensional discrete action space is discretized into multiple tensor grid structures, and the tensor nodes can be expressed as: Among them, G represents the tensor node, i represents the number of the maneuver primitive, and N m represents the total number of maneuver primitives, M i represents the i-th maneuver primitive, j represents the maximum overload amplitude sequence, N n Indicates the total number of maximum overload amplitudes set, n j represents the jth maximum overload amplitude, k represents the distance threshold number for triggering the maneuver primitive, N r Indicates the total number of distance thresholds set, r k represents the kth distance threshold.

5. The interpretable autonomous maneuver penetration method of a UAV according to claim 1, characterized in that: The training, in which the probabilities of different maneuver primitives are updated based on simulation, includes the following sub-steps: S21. At the beginning of each training round, a scene is randomly generated. The continuous action learning automaton selects a tensor node from the current probability density tensor through the inverse cumulative distribution function sampling method. The simulation is performed based on the maneuver primitive, maximum overload amplitude, and distance threshold in the tensor node to obtain the UAV survival probability through simulation. S22. Update the probability density according to the survival probability obtained in this round of training; S23. Repeat the above process for multiple iterations of training to update all probability density tensors. S24. Normalize the entire probability density tensor to form a lookup table.

6. The interpretable autonomous maneuver penetration method of a UAV according to claim 5, characterized in that: In S21, multiple simulations are performed using the selected tensor nodes, and the mean survival probability obtained from multiple simulations is used as the survival probability of the drone in this round of training.

7. The interpretable autonomous maneuver penetration method of a UAV according to claim 6, characterized in that: Initially, the probability density tensor is set to a uniform distribution.

8. The interpretable autonomous maneuver penetration method of a UAV according to claim 5, characterized in that: S22. Update the probability density based on the survival probability obtained in this round of training, expressed as: K j,k =K n (j)K r (k) in, represents the updated probability density tensor, η S Represents the learning rate of reward, R t represents the survival probability obtained in this round of training, η F Represents the learning rate of penalty, P t (M t ,j,k) represents the maneuver primitive M at time t t The probability density at the jth maximum overload amplitude and the kth distance threshold, K j,k , K r (k), K n (j) is the process function, r t Indicates the distance threshold at time t, n t Indicates the maximum overload amplitude at time t, σ r , σ n is the width parameter of the Gaussian kernel function.

9. The interpretable autonomous maneuver penetration method of a UAV according to claim 5, characterized in that: In S23, in each update, the width parameter σ of the Gaussian kernel function r , σ n Gradually become smaller.

10. The interpretable autonomous maneuver penetration method of a UAV according to claim 1, characterized in that: In S3, the UAV also obtains the distance threshold and maximum overload amplitude for triggering the maneuver primitive based on the query lookup table.

Images