METHOD AND DEVICE FOR AUTOMATIC CONTROL OF AN AUTONOMOUS AIRCRAFT

Description

Prior art

[0001] The invention generally relates to aircraft, and in particular to a method and device for automatic guidance of an autonomous aircraft.

[0002] In the field of object surveillance, it is known to use airborne surveillance systems using an aircraft equipped with an imaging unit to monitor objects of interest by taking aerial photographs. The aircraft may be, for example, a multi-rotor drone. The imaging unit may be orientable to photograph objects of interest that may be located in no-fly zones. The aircraft and the imaging unit must be positioned at a precise angle in order to be able to photograph a given object of interest. However, such solutions have significant deployment costs and high energy consumption. Furthermore, the vibrations caused by the motors of the multi-rotor drone and experienced by the imaging unit are high. Such disadvantages considerably shorten the flight time of the drone and significantly degrade the quality of the acquired images.

[0003] Other solutions have been proposed to ensure the automatic guidance of an autonomous aircraft. Generally, the overflight zone considered is not restrictive in the sense that there are no areas prohibited for aircraft overflight. Existing guidance devices apply guidance laws based on optimization problems to determine the shortest path that allows passing through a given number of previously identified objects of interest. However, such guidance laws do not make it possible to anticipate the aircraft's heading to be located in the field of the object of interest to be photographed. A solution of this type is described for example in Patent EP2403757 "Unmanned air vehicle (UAV), control system and method". However, it does not make it possible to anticipate the heading towards an object of interest. Another example of an airborne surveillance system is proposed in document US 2009 / 157233 A1.

[0004] There is thus a need for an improved guidance method and device for autonomous aircraft. General definition of invention

[0005] For this purpose, the present invention relates to a device according to the independent claims. Other embodiments are defined in the dependent claims. Brief description of the figures

[0006] Other characteristics and advantages of the invention will appear with the aid of the following description and the figures in which: [ Fig. 1 ] represents a monitoring system, according to one embodiment of the invention; [ Fig. 2 ] represents an autonomous fixed-wing aircraft, according to one embodiment of the invention; [ Fig. 3 ] illustrates the interaction between the autonomous aircraft and its environment, according to one embodiment of the invention; [ Fig.4 ] illustrates an example of environments in which the autonomous aircraft can operate, according to an embodiment of the invention; [ Fig. 5 ] represents a neural network according to an embodiment of the invention; [ Fig. 6 ] represents steps of a method for optimizing a neural network, according to one embodiment of the invention; [ Fig. 7 ] represents steps of a method for guiding an autonomous aircraft, according to one embodiment of the invention; and [ Fig. 8 ] represents an example of architectures of a guidance device implementing a guidance method, according to an embodiment of the invention. Detailed description

[0007] There figure 1 represents a surveillance system 10 comprising an autonomous aircraft 100 equipped with an imaging unit 101 according to embodiments of the invention. The surveillance system 10 can be configured to monitor a geographic area comprising one or more objects of interest 40 to be photographed.

[0008] The geographic area may include one or more no-fly zones 20 in which overflight by the autonomous aircraft 100 is prohibited. A no-fly zone 20 may be, for example and without limitation, a residential area, an industrial area or an airport. Each no-fly zone 20 may be associated with a three-dimensional (3D) geometry and 3D coordinates defined relative to a given measurement reference frame. The 3D geometry may for example be a polyhedron, a cylinder, or any other 3D shape.

[0009] The geographical area may further comprise one or more safety corridors 30 defining sub-areas that the autonomous aircraft 100 is authorized to fly over. A safety corridor 30 may be configured so as not to cross the no-fly zones 20. A safety corridor 30 may, for example, and in a non-limiting manner, correspond to a road, a river or a beach. A safety corridor 30 may be configured so as to define a trajectory between a departure position and an arrival position. Thus, the autonomous aircraft 100 can reach the arrival position while remaining in the safety corridor 30.

[0010] The geographical area may further comprise one or more objects of interest 40 that can be photographed by the imaging unit 101 equipping the autonomous aircraft 100. An object of interest 40 may belong to a no-fly zone 20 or belong to the safety corridor 30, for example. The objects of interest 40 may be any type of object such as pylons, trees or vehicles. An object of interest 40 may be modeled by a 3D geometry associated with 3D coordinates defined relative to a given measurement reference frame. Furthermore, a marking identifying the type of the object of interest 40 may be associated with each object of interest 40 to be photographed.

[0011] The autonomous aircraft 100 may be, for example, a fixed-wing drone or a multi-rotor drone. Fixed-wing drones have lower vibrations and reduced energy consumption compared to multi-rotor drones. Those skilled in the art will nevertheless understand that the invention is not limited to fixed-wing drones.

[0012] There figure 2 represents an autonomous fixed-wing aircraft 100 according to one embodiment of the invention. The autonomous aircraft 100 may comprise a location unit 103 capable of determining the location of the aircraft in a given measurement frame by returning the three-dimensional coordinates of the aircraft, the location unit 103 being for example of the GPS (Global Positioning System) type. The autonomous aircraft 100 may comprise an inertial unit 104 configured to determine the attitude (roll, pitch and heading angles) of the autonomous aircraft 100 in the given measurement frame. The autonomous aircraft 100 may further comprise a speed sensor and an orientation sensor (not shown in the figure). figure 2 ) configured respectively to determine the instantaneous speed and the orientation of the autonomous aircraft 100 in the given measurement reference frame.

[0013] Advantageously, the autonomous aircraft 100 may further comprise one or more position sensors 105 arranged on its outer surface, at different locations, each position sensor 105 being configured to detect obstacles likely to cause a collision involving the autonomous aircraft 100 and to determine the distance separating the autonomous aircraft 100 from each detected obstacle. The position sensors 105 may further be configured to identify the nature of the detected obstacles. In one embodiment, the position sensors 105 may for example be arranged uniformly and symmetrically around the nose of the autonomous aircraft. The number of position sensors 105 may for example be odd by positioning one of them along the longitudinal axis of the autonomous aircraft. Furthermore, the position sensors 105 may be configured to return a negative value, for example -1, in the event of no obstacle detection.

[0014] The autonomous aircraft 100 may further comprise an object of interest sensor 40 (not shown in the figure 2 ) configured to detect objects of interest 40 and to determine the relative position of each detected object of interest 40 with respect to the instantaneous position of the autonomous aircraft 100. Such a relative position may be represented by an angle and a distance separating the autonomous aircraft 100 from the detected object of interest 40. Alternatively, the coordinates in a given measurement frame of the objects of interest 40 may be saved in advance by the object of interest sensor 40. In this case, the object of interest sensor 40 may be configured to determine the relative position of each object of interest 40 from the coordinates of the object of interest 40 and the instantaneous position of the autonomous aircraft 100 as provided by the location unit 103 described above.

[0015] The autonomous aircraft 100 may further comprise an automatic pilot 102 configured to execute a guidance law and to implement received control instructions, according to the executed guidance law, by actuation of elements of the autonomous aircraft 100 such as its engine and its control surfaces which act respectively on the speed and on the attitude of the autonomous aircraft 100.

[0016] The autonomous aircraft 100 may also comprise an imaging unit 101 which may be rigidly fixed to the aircraft. The imaging unit 101 may be associated with a set of opto-geometric parameters such as the focal length, the field angle and the orientation which may be fixed over time. In this case, the imaging unit 101 may be oriented forward, along the longitudinal axis of the autonomous aircraft 100. Alternatively, one or more opto-geometric parameters of the imaging unit 101 may be adjusted over time. For example, the orientation and the focal length of the imaging unit 101 may be adjusted over time. Furthermore, the imaging unit 101 may be configured to receive shooting instructions intended to activate the photographic shooting of an object of interest 40 located in its field of vision.The photographic shooting instructions may include, for an imaging unit 101 with adjustable opto-geometric parameters, the values ​​of such adjustable parameters.

[0017] The autonomous aircraft 100 may further comprise a storage unit (not shown in the figure 2 ) configured to permanently save, i.e. non-volatile, data of a geographical nature linked to the position and geometric shape of the elements comprising the no-fly zones 20, the safety corridor 30 and / or the objects of interest 40 to be photographed.

[0018] Advantageously, the autonomous aircraft 100 may further comprise a guidance device 106 configured to communicate with elements of the autonomous aircraft 100 comprising the autopilot 102, the storage unit, the imaging unit 101 and the sensors equipping the autonomous aircraft 100 (location unit 103, inertial unit 104, speed sensor, orientation sensor, position sensors 105, object of interest sensor 40, etc.). The guidance device 106 may be configured to determine, at a given instant during the flight of the autonomous aircraft 100, a current state of the autonomous aircraft 100 which may be described in relation to the external elements comprising the safety corridor 30, the no-fly zones 20, the obstacles likely to cause collisions, objects of interest 40 to be photographed, etc. Such external elements form the environment 200 of the autonomous aircraft 100.Advantageously, the guidance device 106 can further be configured to govern the interaction between the autonomous aircraft 100 and its environment 200 by acting on the behavior (attitude and speed) of the autonomous aircraft 100.

[0019] There figure 3 illustrates the interaction between the autonomous aircraft 100 and its environment 200 as governed by the guidance device 106, according to one embodiment of the invention. In this case, the guidance device 106 may further be configured to determine, at a given instant during the flight of the autonomous aircraft 100, an action to be executed from a set of possible actions, the action to be executed possibly being arbitrary or dependent on the current state of the autonomous aircraft 100. Such an action to be executed may for example consist of changing the heading of the autonomous aircraft 100 by 10 degrees and / or increasing the speed by 30%. Such a change in behavior (attitude and speed) is perceptible by the various sensors equipping the autonomous aircraft 100 which allow the guidance device 106 to determine a new state of the aircraft in its environment, as well as an associated reward, also called a reward parameter.The reward may for example be real-valued, a negative reward being representative of a “punishment”. For example, the reward may be equal to 0.5 or -0.5 depending on whether the new state corresponds to an aircraft remaining in the safety corridor 30 or corresponds to an aircraft leaving the safety corridor 30. In another example, the reward may be equal to 1 if the new state allows the autonomous aircraft 100 to photograph an object of interest 40 while remaining in the safety corridor 30 and may be equal to 2 if the new state allows the autonomous aircraft 100 to reach the arrival position while remaining in the safety corridor 30. The action executed by the autonomous aircraft 100 may for example be described as “optimal” if it corresponds to a maximum reward.The interaction between the autonomous aircraft 100 and its environment 200 as governed by the guidance device 106 can be iterative where each iteration is representative of a stage of the flight of the autonomous aircraft 100. The flight of the autonomous aircraft 100 can therefore be likened to an episode corresponding to a series of stages where each stage is represented by a state, an action and a reward.

[0020] There figure 4 illustrates, simply as an example, a representation of the environment 200 in which the autonomous aircraft 100 can operate. The columns of the matrix of the figure 4 represent stages of an episode, i.e. of a flight of the autonomous aircraft 100 between a starting position and an arrival position. The boxes of the matrix of the figure 4 represent the states in which the autonomous aircraft 100 can find itself. For the sake of simplification, only four states are considered in the example of the figure 4 and include: a first state corresponding to an aircraft falling outside the safety corridor 30, the first state being represented by the multiplication sign “X” and associated with a reward equal to -0.5, for example; a second state corresponding to an autonomous aircraft 100 inside the safety corridor 30, the second state being represented by a circle “O” and associated with a reward equal to 0.5, for example; a third state corresponding to an autonomous aircraft 100 inside the safety corridor 30 and photographing an object of interest 40, the third state being represented by the letter “V” and associated with a reward equal to 0.5, for example; a fourth state corresponding to an autonomous aircraft 100 reaching the arrival position, the fourth state being represented by a triangle and associated with a reward equal to 2, for example;

[0021] In the simplified example of the figure 4 , the autonomous aircraft 100 may have a set of three possible actions consisting of moving left, right or forward.

[0022] In embodiments of the invention, the guidance device 106 of the autonomous aircraft 100 can be configured to determine an optimal strategy, also called optimal policy, making it possible to choose, in relation to each current state of the autonomous aircraft 100, an optimal action to be executed.

[0023] Advantageously, the determination of the optimal strategy can be carried out during a learning phase prior to the actual flight phase of the autonomous aircraft 100. The learning phase can be carried out using fictitious flights of the autonomous aircraft carried out in environments 200 which are also fictitious and which may not be identical to the actual environment 200 in which the autonomous aircraft will operate, each fictitious environment 200 being representative of a scenario.

[0024] In embodiments of the invention, the strategy may be represented by an action-state value function, denoted Q[s, a], which determines the potential gain, also called the action value, i.e. the weighted sum of the rewards of each future step from the current state, provided by performing a certain action a in a certain state s. When the value function Q[s, a] is learned, i.e. optimized, by the guidance device 106 of the autonomous aircraft 100, the optimal strategy may be obtained by selecting the action with maximum value for each state, i.e. by selecting the action a which maximizes the value Q[s, a] when the autonomous aircraft 100 is in state s. Furthermore, the action-state value function may be represented by a table, also called a Q-table, in which each possible state is associated with a row and each possible action is associated with a column.In the Q-table, the value of each bin represents the value of the state-action value function for the associated state-action pair. The elements of a Q-table can be initialized to arbitrary values ​​and then optimized using a given reinforcement learning algorithm.

[0025] Advantageously, the optimization of an action-state value function represented by a Q-table can be carried out using a reinforcement learning algorithm of the “Q-Learning” type. In this case, the definition of the action-state value function can be updated after each step of a fictitious flight (episode) of the autonomous aircraft 100 according to the following relationship: Q s a = 1 − α Q s a + α r + γ max a Q s ′ , a where s and s' represent the new state and the current state respectively, a represents the action executed, r represents the reward received, α is a learning factor between 0 and 1, and γ is an update factor. Relation (1) therefore makes it possible to replace the value of a box in the Q-table with another value depending on the reward received.

[0026] Furthermore, the choice of the action to be executed a can be carried out according to an approach specified by the strategy Q[s, a]. Such an approach can for example be of the ε-greedy type (ε ∈ [0, 1]) according to which the autonomous aircraft 100 randomly draws a number between 0 and 1. If this number is less than the coefficient ε, the aircraft will explore its environment 200 by choosing an arbitrary action; otherwise it will exploit the information from the action-state value function. Advantageously, the value of ε can be constant for all the stages of an episode and can decrease when moving from one episode to another episode. For example, the update of the action-state value function in relation to the first fictitious flight of the autonomous aircraft 100 can be carried out at a coefficient ε equal to 1. This corresponds to an exploration phase during which the autonomous aircraft 100 chooses random actions as it does not know its environment 200.Furthermore, when the value of the coefficient ε is close to zero, for example substantially equal to 0.05, the autonomous aircraft 100 can use the action-state value function as updated to choose the optimal action having the maximum potential gain. This corresponds to an operating phase during which the autonomous aircraft 100 exploits the knowledge about its environment 200.

[0027] Alternatively, the optimization of an action-state value function represented by a Q-table can be performed using a reinforcement learning algorithm of the “Double Q-learning” type. Such a variant of the “Q-learning” algorithm makes it possible to accelerate the convergence of the strategy by implementing two action-state value functions whose updating is carried out crosswise according to the following two relations: Q A s a = Q A s a + α r + γQ B s ′ , arg max a Q A s ′ , a − Q A s a Q B s a = Q B s a + α r + γQ A s ′ , arg max a Q B s ′ , a − Q B s a

[0028] In other embodiments of the invention, the number of states in which the autonomous aircraft 100 can find itself can be very high, or even infinite. Indeed, each possible combination of the input information provided by the sensors equipping the autonomous aircraft 100 defines a different current state. It is therefore clear that a representation of the action-state value function according to a matrix format as described above can be complex in terms of computing power and in terms of memory resources. Advantageously, the action-state value function can be represented, in an approximate manner, by a deep neural network 500 in order to address the problem of the large size of the state space of the autonomous aircraft 100.In this case, the neural network 500 may be formed of several layers each comprising one or more neurons where each neuron of a given layer is connected to all the neurons of the previous layer and to all the neurons of the following layer, as illustrated in the . figure 5 , the connection between two neurons being associated with a given weight. The neural network 500 comprises an input layer 501 comprising several neurons where each neuron receives input information, also called an input parameter. The input information can be directly provided by a sensor equipping the autonomous aircraft 100 or calculated from the information provided by one or more sensors and / or provided by the storage unit equipping the autonomous aircraft 100. The neural network 500 can comprise several intermediate layers 502, each of them being formed of one or more neurons. The neural network 500 further comprises an output layer 503 comprising several neurons where each neuron corresponds to a possible action that the autonomous aircraft 100 can undertake. The number of possible actions can advantageously depend on the guidance law implemented.

[0029] Furthermore, the optimization of an action-state value function represented by a neural network 500 amounts to optimizing the parameters of the neural network 500, i.e. the weights of the neural network 500, using several fictitious flights of the autonomous aircraft 100.

[0030] Advantageously, the optimization of an action-state value function represented by a neural network 500 can be carried out using a reinforcement machine learning algorithm of the “Deep Q-Learning” type. According to such a machine learning algorithm, the parameters of the neural network 500, represented by a vector denoted θ, are initialized with random weights before being updated after each stage of a fictitious flight by stochastic gradient descent using the following cost function: L θ = Q s , a ; θ − r + γ max a Q s ' , a ; θ 2 (4) where Q[s, a; θ] represents the value of Q[s, a] estimated with the 500 neural network and (r + γ max a Q[s', a; θ]) represents the target value. In relation (4), the same 500 neural network is used to select the action and to calculate the target value.

[0031] Alternatively, the optimization of an action-state value function represented by a 500 neural network can be performed using a reinforcement learning machine learning algorithm of the “Double Deep Q-Learning” type. Such a variant of the “Deep Q-Learning” algorithm has the advantage of separating the selection of the action that maximizes the value of the function Q[s, a] from its evaluation using two separate neural networks: a first neural network 500 for selecting the action to be executed, denoted “Q” network; a second neural network 500 for evaluating the action executed, denoted “Q-target” network.

[0032] There figure 6 represents steps implemented in the optimization of an action-state value function represented by a neural network 500 using a “Double Deep Q-Learning” type reinforcement machine learning algorithm, according to one embodiment of the invention. Such steps are described in relation to a step of a fictitious flight of the autonomous aircraft 100 and are implemented recursively after an initialization step, each iteration being associated with a step of a fictitious flight of the autonomous aircraft 100.

[0033] The initialization step (not shown in the figure 6 ) consists of initializing a recovery memory whose capacity is previously defined. The initialization step also consists of initializing the parameters of the two neural networks, “Q” network and “Q-target” network, with random weights, θ INI Q for the “Q” network and θ INI Q − cible for the “Q-target” network.

[0034] Advantageously, an observation step (not shown on the figure 6 ) can be implemented at the start of each new episode before the implementation of the first iteration of steps 601 to 605 relating to the first step of the new episode. The observation step consists of observing the initial state of the autonomous aircraft 100.

[0035] In step 601, the autonomous aircraft 100 can select and implement an action according to a given approach which can be of the ε-greedy type as described above.

[0036] In step 602, the autonomous aircraft 100 can determine the reward and the new state from the information provided by the sensors and / or by the storage unit equipping the autonomous aircraft 100. Step 602 can further consist of storing in the recovery memory the transition represented by a set formed by the current state, the action executed, the reward received and the new state.

[0037] At step 603, the autonomous aircraft 100 may randomly select a set representative of a transition from the sets stored in the resume memory.

[0038] At step 604, the autonomous aircraft 100 may calculate the action-state value function Q s a θ ACTUEL Q using the 500 “Q” neural network. Further, the autonomous aircraft 100 can calculate ( r + γ max a Q s ′ , a ; θ ACTUEL Q − cible ) using the 500 “Q-target” neural network.

[0039] In step 605, the autonomous aircraft 100 can update the parameters of the neural network 500 by stochastic gradient descent using the following cost function: L θ = Q s a θ ACTUEL Q − r + γ max a Q s ′ , a ; θ ACTUEL Q − cible 2

[0040] Advantageously, a cloning step (not shown on the figure 6 ) can be implemented on a regular basis, for example after the end of a given episode. The cloning step can consist of using the parameters of the 500 “Q” neural network to update the parameters of the 500 “Q-target” neural network.

[0041] Generally, any reinforcement machine learning algorithm may be used to optimize the action-state value function implemented in the guidance device 106, the value function being able to be represented by a Q-table or advantageously represented by a neural network 500. Examples of reinforcement machine learning algorithms include, by way of non-limiting example, the “Monte Carlo” method, the “Temporal-Difference Learning” algorithm or the “Exhaustive Search” algorithm.

[0042] In one embodiment of the invention, a complexity parameter may be associated with each fictitious flight of the autonomous aircraft 100, the fictitious flight being carried out in a fictitious environment 200. The complexity parameter may for example represent the number of turns in the safety corridor 30 associated with the fictitious environment 200. Advantageously, the selection of the fictitious environments 200 used in the learning phase may be done according to an increasing order of the complexity parameters, that is to say starting with the environments 200 having low complexity parameters.

[0043] In another non-limiting exemplary embodiment, the learning phase can be initiated by a fictitious flight associated with a complexity parameter defined from one or more turns present in the safety corridor 30 and / or one or more objects of interest 40. The learning phase can then continue progressively by increasing in complexity (in increasing order of complexity parameters associated with the fictitious flights), by adding other turns, in the safety corridor 30, other objects of interest 40, and generally other elements defining a guidance constraint for the aircraft in the safety corridor 30.

[0044] In another embodiment of the invention, the strategy can be directly represented by a function determining for each current state an optimal action to be executed. Such a function can for example be represented by a neural network 500 as described above. Advantageously, the neural network 500 can in this case be optimized using a reinforcement machine learning algorithm of the “Proxima Policy Optimization” type. Advantageously, the neural network 500 can in this case be a recurrent neural network 500.

[0045] There figure 7 represents steps of a method for guiding an autonomous aircraft 100 according to one embodiment of the invention. Such steps are implemented by the guidance device 106 of an autonomous aircraft 100 implementing a neural network 500 determined and optimized during a prior learning phase according to one of the embodiments of the invention described above. The steps of the figure 7 are implemented at a given moment in the flight of the autonomous aircraft 100.

[0046] In step 701, the guidance device 106 can determine the current state of the autonomous aircraft 100 from the information provided by the sensors equipping the autonomous aircraft 100 and / or by its storage unit. Such a current state can for example comprise the position, speed and attitude of the autonomous aircraft 100, the relative position of an object of interest 40 to be photographed, the distances separating the autonomous aircraft 100 from the detected obstacles, etc.

[0047] In step 702, the guidance device 106 may determine an optimal action to be performed using the previously determined and optimized neural network 500. The neural network 500 uses as input the current state of the autonomous aircraft 100 as determined in step 701 and provides an optimal action that the autonomous aircraft 100 can take.

[0048] In step 703, the guidance device 106 can determine several control instructions intended for the automatic pilot 102 equipping the autonomous aircraft 100 from the optimal action obtained in step 702. Thus, the optimal action is transformed into a plurality of control instructions. Step 703 can be carried out according to the guidance law implemented in the automatic pilot 102.

[0049] In step 704, the guidance device 106 can transmit to the automatic pilot 102 the control instructions in order to modify the behavior of the autonomous aircraft 100 in terms of speed and / or in terms of attitude, which provides a new state of the autonomous aircraft 100.

[0050] In embodiments of the invention, steps 702 and 703 can be combined into a single step so that the neural network 500 determines, from the current state received, the control instructions intended for the automatic pilot 102 equipping the autonomous aircraft 100.

[0051] Advantageously, the guidance method may further comprise an additional step (not shown in the figure 7 ) consisting of transmitting to the imaging unit 101 a photographic shooting instruction if the new state of the autonomous aircraft 100 allows one or more objects of interest 40 to be photographed.

[0052] In one embodiment of the invention, the autonomous aircraft 100 is of the fixed-wing autonomous aircraft 100 type and the guidance law implemented to modify the behavior of the autonomous aircraft 100 can be chosen from the following laws: a law for capturing and maintaining a heading, by action on the control surfaces of the ailerons of the autonomous aircraft 100; a law for capturing and maintaining a course by action on the control surfaces of the ailerons of the autonomous aircraft 100; a law for capturing and maintaining altitude by action on the thrust of the engine or on the elevators of the autonomous aircraft 100; a law for capturing and maintaining speed by action on the thrust of the engine of the autonomous aircraft 100.

[0053] There figure 8 represents the architecture of a guidance device 106 implemented in an autonomous aircraft 100 according to an embodiment of the invention. The guidance device 106 may comprise: a first unit 801 capable of determining the current state of the autonomous aircraft 100 from the information provided by the sensors used and / or the information provided by the storage unit equipping the autonomous aircraft 100; a second unit 802 capable of determining an optimal action to be executed using a previously determined and optimized neural network 500; a third unit 803 capable of determining several control instructions intended for the automatic pilot 102 from the optimal action to be executed; a fourth unit 804 capable of transmitting to the automatic pilot 102 the control instructions in order to modify the behavior of the autonomous aircraft 100.

[0054] The guidance device 106 is advantageously configured to implement a preliminary learning phase to optimize the neural network 500 from a plurality of fictitious flights of the aircraft, each of the fictitious flights comprising a succession of fictitious steps.

[0055] The guidance device 106 may further comprise a fifth unit 805 capable of transmitting photographic shooting instructions to the imaging unit 101.

[0056] The embodiments of the invention can be advantageously applied to an autonomous aircraft 100 of the fixed-wing drone type equipped with an imaging unit 101 with fixed opto-geometric parameters over time. This makes it possible to increase the flight duration and to cover larger geographical areas than the solutions of the state of the art.

[0057] Those skilled in the art will understand that the guidance method according to the embodiments may be implemented in various ways by hardware, software, or a combination of hardware and software, including in the form of program code that may be distributed as a program product, in various forms. In particular, the program code may be distributed using computer-readable media, which may include computer-readable storage media and communication media. The methods described in the present disclosure may be implemented in particular as computer program instructions executable by one or more processors in a computer computing device. These computer program instructions may also be stored in a computer-readable medium.

[0058] The invention is not limited to the embodiments described above as a non-limiting example. It encompasses all variant embodiments that may be envisaged by those skilled in the art.

Claims

1. A method for guiding an autonomous aircraft (100), the aircraft comprising an automatic pilot (102), a plurality of sensors and an imaging unit (101), the aircraft being configured to fly over a geographic zone comprising overflight prohibited zones (20) and a safety corridor (30) that does not pass through said overflight prohibited zones (20), said imaging unit (101) being able to photograph one or more objects of interest (40) located in said geographic zone, the guidance method comprises a phase of real flight of the autonomous aircraft (100) between a position of departure and a predefined position of arrival that are associated with said safety corridor (30), by using a given guidance law, the guidance method comprising the following steps, implemented at a given instant of said real flight phase: - determining (701) a current state of said autonomous aircraft (100) as a function of a plurality of input parameters comprising input parameters provided by said plurality of sensors; - determining (702) an optimum action to be executed by using a neural network (500) receiving said current state; - determining (703) a plurality of control instructions compatible with said guidance law based on said optimum action to be executed; - transmitting (704) to said automatic pilot (102) said plurality of control instructions, which provides a new state of said autonomous aircraft (100); and characterised in that the method comprises a preliminary learning phase of optimising said neutral network (500) based on a plurality of fictitious flights of said aircraft, each of said fictitious flights comprising a succession of fictitious steps and being associated with a complexity parameter, said fictitious flights being used in said preliminary learning phase according to an ascending order of said complexity parameters.

2. The guidance method according to claim 1, characterised in that it comprises an additional step of transmitting to said imaging unit (101) photographic exposure instructions, if said new state makes it possible to photograph at least one object of interest (40).

3. The guidance method according to claim 1 or 2, characterised in that said neural network (500) represents a value function, said value function associating one or more actions with each current state of said autonomous aircraft (100), each of said actions being associated with an action value representative of an expected sum of reward parameters, said optimum action to be executed being associated with a maximum action value.

4. The guidance method according to claim 1 or 2, characterised in that said neural network (500) represents a strategy associating an optimum action to be executed with each current state of said autonomous aircraft (100).

5. The guidance method according to claim 3 or 4, characterised in that the optimisation of said neutral network (500) based on said plurality of fictitious flights is determined by using a reinforcement machine learning algorithm.

6. The guidance method according to one of the preceding claims, characterised in that said neural network (500) comprises an input layer (501) formed by a plurality of input neurons, at least one of said input neurons receiving a piece of information provided by one of said sensors.

7. The guidance method according to claim 3, characterised in that said reward parameters are maximal if the new state corresponds to: - a state indicating that the autonomous aircraft (100) remains in the safety corridor (30); - a state indicating that the imaging unit (101) photographs at least one object of interest (40); - a state in which the autonomous aircraft (100) reaches a predefined position of arrival, and in that said reward parameters are minimal if the new state corresponds to a state indicating an exit of the autonomous aircraft (100) from the safety corridor (30).

8. The guidance method according to one of the preceding claims, characterised in that said guidance law is a law from among: - a heading capture and maintenance law, by action on control surfaces of the ailerons of the autonomous aircraft (100); - a route capture and maintenance law, by action on control surfaces of the ailerons of the autonomous aircraft (100); - an altitude capture and maintenance law, by action on the thrust of the engine or on elevators of the autonomous aircraft (100); - a speed capture and maintenance law, by action on the thrust of the engine of the autonomous aircraft (100).

9. A guidance device (106) implemented in an autonomous aircraft (100), the aircraft comprising an automatic pilot (102), a plurality of sensors and an imaging unit (101), the aircraft being configured to fly over a geographic zone comprising overflight prohibited zones (20) and a safety corridor (30) that does not pass through said overflight prohibited zones (20), said imaging unit (101) being able to photograph one or more objects of interest (40) located in said geographic zone, the guidance device (106) comprises: - a first unit (801) capable of determining a current state of said autonomous aircraft (100) as a function of a plurality of input parameters comprising input parameters provided by said plurality of sensors; - a second unit (802) capable of determining an optimum action to be executed by using a neural network (500) receiving said current state; - a third unit (803) capable of determining a plurality of control instructions compatible with said guidance law based on said optimum action to be executed; - a fourth unit (804) capable of transmitting to said automatic pilot (102) said plurality of control instructions, which provides a new state of said autonomous aircraft (100), characterised in that the guidance device (106) is configured to implement a preliminary learning phase to optimise said neutral network (500) based on a plurality of fictitious flights of said aircraft, each of said fictitious flights comprising a succession of fictitious steps and being associated with a complexity parameter, said fictitious flights being used in said preliminary learning phase according to an ascending order of said complexity parameters.

10. The guidance device according to claim 9, characterised in that it comprises a fifth unit (805) capable of transmitting photographic exposure instructions to the imaging unit (101).

11. An autonomous aircraft (100), characterised in that it comprises a guidance device (106) configured according to one of claims 9 and 10.