Autonomous watercraft and method for operating an autonomous watercraft

Abstract

The invention relates to a method for operating an autonomous watercraft (20) with a drive and control unit (28) for maneuvering the autonomous watercraft (20), comprising the following steps: a) detecting environment information of the autonomous watercraft by means of a sensor system (24) comprising a sensor or a plurality of sensors, b) outputting one sensor signal (26) per sensor (24'), corresponding to the environment information; c) actuating the drive and control unit (28) in such a way that the watercraft (20) travels along a preplanned route; d) detecting an event on the basis of the sensor signal or the sensor signals and determining a risk of the event for the watercraft (20); e) actuating the drive and control unit (28) in such a way that the watercraft performs an evasive maneuver in order to avoid or mitigate the event, if the risk exceeds a threshold value; f) training an algorithm by means of machine learning for controlling the autonomous watercraft by means of training data on the basis of the sensor signals and on the basis of control signals for actuating the drive and control unit.

Description

[0001] Autonomous watercraft and procedures for operating an autonomous watercraft

[0002] Description

[0003] The invention relates to training a model to control an autonomous watercraft based on the existing deterministic autonomy of an autonomous watercraft.

[0004] Autonomous watercraft can move completely independently, i.e., without remote control. Currently, autonomous watercraft are controlled by a system, primarily deterministic, based on general maritime rules. Remote control may still be possible as a supplement. It is known that autonomous watercraft can follow a pre-planned route. However, the challenge lies in the watercraft's ability to react independently to unforeseen and unpredictable events that occurred during route planning.

[0005] The object of the present invention is therefore to create an improved concept for autonomous watercraft.

[0006] The problem is solved by the subject matter of the independent patent claims. Further advantageous embodiments are the subject matter of the dependent patent claims.

[0007] Exemplary embodiments show an autonomous watercraft. The watercraft comprises a propulsion and control unit, a sensor system, and a data processing unit. The propulsion and control unit can maneuver the autonomous watercraft. That is, the speed, direction of travel, and optionally the diving depth of the autonomous watercraft can be set using the propulsion and control unit. The design of the propulsion and control unit can include a common component for propulsion and control or separate components for propulsion and control. A propulsion and control unit with common components includes, for example, one or more propellers or one or more nozzles for propulsion, whereby control is achieved by pivoting the propeller(s) or nozzles or by varying the drive power of the propeller(s) or nozzles relative to each other. A propulsion and control unit with separate components includes, for example,In addition to a propulsion system, a rudder may be used for steering. Furthermore, the propulsion and steering unit can include one or more trim systems for adjusting lift, heel, and / or pitch. In other words, the propulsion and steering unit can also influence the attitude of the autonomous watercraft, for example, by means of a trim system.

[0008] The sensor system comprises one or typically multiple sensors. The sensor(s) detect environmental information from the autonomous underwater vehicle and output sensor signals corresponding to that environmental information. For example, a camera records (especially continuously) images of the surroundings, while an infrared sensor (especially continuously) detects heat signatures of the environment, a sonar (especially continuously) detects underwater sound, an AIS receiver (especially continuously) receives position data from ships, and a radar (especially continuously) transmits radio waves and receives their echoes. Additional sensors can also be used. The sensors can form the sensor system as a single sensor or in any combination. The sensor system can also include sensors that monitor the internal status of the autonomous underwater vehicle.

[0009] The data processing unit can control the propulsion and steering unit, particularly via control signals, so that the vessel travels along a pre-planned route. The data processing unit represents one option for controlling the autonomous vessel during regular operation. However, at times, a sensor from the sensor system, particularly via suitable interfaces, can take over regular control, or an external control center can temporarily assume regular control. Specifically, the data processing unit can control the vessel using a set of rules based on sensor data. Furthermore, the data processing unit can detect an event based on the sensor signal(s) and determine the risk posed by that event to the vessel. The event could be an environmental phenomenon such as a thunderstorm or a storm front.Furthermore, the event can also be a (moving or stationary) obstacle, such as a moving or anchored vessel. The risk in this case is, for example, a collision or potential damage to the autonomous vessel caused by the environmental event. If the risk exceeds a threshold, the propulsion and control unit is controlled by the data processing unit via control signals in such a way that the vessel executes an evasive maneuver to avoid or at least mitigate the event. The evasive maneuver can include changing the speed and / or the direction of travel, i.e., changing the current route of the autonomous vessel. The evasive maneuver also optionally includes returning to the pre-planned route as soon as this is reasonable and possible. Alternatively, the data processing unit can calculate an alternative route to the pre-planned route to reach the destination.

[0010] Furthermore, the data processing unit is equipped to train a mathematical model for controlling the autonomous watercraft using machine learning. This model is based on training data derived from sensor signals and control signals for the propulsion and control unit. Specifically, the sensor signals can represent the input data for the model, and the control signals can represent the model's output.

[0011] The idea is therefore to train a model, i.e., an artificial intelligence, using machine learning based on a functioning, deterministic system. This approach has at least one significant advantage: it eliminates the need for manual training. Instead, the model is trained during the regular operation of the autonomous watercraft. The training takes place directly on the autonomous watercraft. No additional resources beyond the necessary computing power of the data processing unit, particularly human labor, are required for training. A model describes the architecture of an algorithm. Most common, well-known algorithms are suitable, especially artificial neural networks, Bayesian classifiers, support vector machines (SVMs), and linear regression, but deep learning approaches are preferred.Artificial neural networks based on transformer technology and / or convolutional artificial neural networks. That is, the model is preferably implemented as an artificial neural network. The artificial neural network can, for example, be based on transformer technology and / or be or comprise a convolutional artificial neural network.

[0012] In particular, it has been shown that no special algorithms are needed to create the model. It is sufficient to train commercially available algorithms with suitable data. Such algorithms are also known as COTS (Components Off-The-Shelf). It has also been found that the structure—for example, the number of neurons or the number of layers in artificial neural networks—is irrelevant for training the model, as long as a minimum level of complexity is achieved. Models that perform well in public benchmarks are particularly suitable. For example, the mathematical model could be a neural network in the form of a multimodal transformer model.For training, sensor data, such as the current course and / or images and / or time series, for example from sonar data, can be used as input values ​​for both training and operation. Different sub-models, suitable for processing the specific type of sensor data, can be used to process the sensor data. The mathematical model then outputs control data for the vessel. For training, this control data can be compared with the control data generated by the data processing unit. This enables supervised learning of the mathematical model during regular vessel operation. The use of a multimodal transformer model is described, for example, in the following source: "ZHANG, Yiyuan, et al. Meta-transformer: A unified framework for multimodal learning. arXiv preprint arXiv:2307. 10802, 2023."“In addition or as an alternative, a large language model (LLM) can also be used. Examples of LLMs are Gemma or Llama 3.2 vision. The Llama model is described, for example, in the following source: “GRATTAFIORI, Aaron, et al. The Hama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024.”.”

[0013] In exemplary embodiments, the data processing unit is configured to determine the model's quality after it has been trained with a large amount of training data, before further training using actual sensor signals. This means that the training can be interrupted for, for example, a specific number of decisions, such as generating different control signals or taking a specific number of sensor samples. During this time, the model receives the sensor signals and generates its results based on them. These results are then compared, for example, with the control signals to determine the model's quality.

[0014] If the model's quality exceeds a threshold, the data processing unit can then control the propulsion and steering unit based on the model's results. A suitable threshold could be, for example, a certain percentage of the model's results that are identical or similar to the control signals. A result is considered similar to the control signals, for instance, if the vessel performs the same maneuver over a certain period of time. Alternatively or additionally, a maximum deviation in the deflection of the steering unit, such as a rudder, can be defined.

[0015] Furthermore, the data processing unit may contain a database of general navigation rules. The data processing unit can compare the model's output with this database to ensure the autonomous vessel operates in accordance with these rules. Specifically, if the comparison reveals a violation, the unit can initiate control signals from the propulsion and steering system. In other words, a model output can be reviewed or discarded if it does not comply with the navigation rules. In this case, the control signals can be used for steering, at least temporarily. Alternatively, a warning can be issued, prompting manual control, i.e., remote control, of the autonomous vessel.

[0016] In further embodiments, the data processing unit includes an interface for receiving external commands. The data processing unit can replace the control signals generated by adhering to the pre-planned route or performing an evasive maneuver with these external commands in order to control the propulsion and steering unit based on the external commands. This means that remote control of the autonomous watercraft is possible using the external commands. For training the model, the data processing unit replaces the control signals with the external commands. Consequently, the control signals are considered incorrect when the autonomous watercraft is manually intervened, thus necessitating the intervention in the first place. Therefore, it is possible for the model, which also learns from the external commands, to become more proficient than the control signals themselves.In this respect, the threshold of the model, at which the result of the model is used to control the drive and control unit, may be exceeded even if the model is improved compared to the original control signals.

[0017] External commands can include, among other things, movement instructions (e.g., speed, direction of travel, and / or diving depth) for remote control of the vessel, maneuver instructions, sensor settings, or waypoints to be navigated to. These external commands can be generated by a data processing system integrated into the vessel that is not inherently assigned to the data processing unit. Alternatively, such commands can also be transmitted via communication systems, e.g., from a control center, and made accessible to the data processing unit. The data processing unit is designed to replace the commands generated by adhering to the pre-planned route with the external commands in order to control the propulsion and steering unit based on these external commands.Depending on the type of external commands, the data processing unit can completely or partially suspend monitoring of adherence to the predetermined route or override specific commands with the corresponding external commands. For remote control, the autonomous watercraft can transmit its sensor data to the control center. Communication with an autonomous surface vessel can be established using electromagnetic waves, such as radio communication. For communication with an autonomous underwater vehicle, the use of a cable, particularly an optical fiber, or acoustic waves (e.g., an acoustic modem) is recommended.

[0018] In some embodiments, the data processing unit is further configured to continuously monitor the sensor signals for events, even during remote control, and, if the risk of the event to the vessel exceeds the threshold, to control the propulsion and steering unit in such a way that the vessel performs an evasive maneuver to avoid a collision. That is, the data processing unit continues to monitor the autonomous vessel even when it is not currently in control and takes over control if exceptional events, such as a potential collision or an approaching storm, are likely to occur.

[0019] In other words, it may be possible for the external command to be overridden by the current control signal in an emergency situation. This can happen, for example, if the vessel is on a collision course and no evasive maneuver is initiated by the external commands. Before a collision occurs, the autonomous vessel can then regain control. The data processing unit is then trained to use the control signal, and not the external command, to train the model. Control can also be taken over if a signal loss occurs during remote control by the control center and the connection is interrupted. In this case, the vessel can take steps to re-establish the connection or alternatively wait for it to be reconnected. These options can optionally be disabled by the control center.Communication between the autonomous watercraft and the control center, particularly the transmission of video images, can be carried out via Ethernet, for example. The system can automatically select the best route among various available connection types. Typical connection types include WiFi, IP over VHF, LTE, or SATcom.

[0020] Preferably, during remote control by the control center, the data processing unit can detect an event based on the sensor signal(s) and determine the risk of the event to the vessel. If the risk exceeds a threshold, the data processing unit can control the propulsion and steering unit in such a way that the vessel performs an evasive maneuver to avoid or mitigate the event. This means that even during remote control by the control center, the data processing unit can take over control if there is a danger to the autonomous vessel.

[0021] In some implementations, however, the data processing unit at the interface may receive information to interrupt the model training. For example, an operator intervention, such as pressing a button, can signal that their input should not be used to train the model. This might occur, for instance, when the autonomous underwater vehicle has completed its mission and is being operated manually, or when a deviation from the mission is required for other reasons.

[0022] Further examples demonstrate that the data processing unit is designed to train multiple different models simultaneously. This means, for example, that various machine learning methods can be used, such as artificial neural networks, linear regression, Bayesian classifiers, hidden Markov models, etc. It can also utilize different topologies of a single machine learning method.

[0023] Depending on the quality of any number of models from a plurality of different models, or after a predefined training period based on a decision criterion, the data processing unit can control the propulsion and steering unit based on the results of the models from the plurality of models with selected outcomes. For example, a majority decision can be applied as a decision criterion to control the propulsion and steering unit according to a selection of the results based on the majority decision. One possible implementation is, for example, to use all results that steer the watercraft in the correct direction and calculate a mean or median of these results to obtain the actual control signal for the propulsion and steering unit.

[0024] As an alternative to majority voting, it is also possible to select the best model from the majority of different models as the decision criterion and to control the drive and control unit based on the results of the best model. The best model can be determined once or at specific intervals in advance based on the quality of all models.

[0025] Exemplary embodiments show that the sensor system includes at least one optical sensor. The optical sensor is, for example, a camera or a lidar sensor (Lidar: Light detection and ranging).

[0026] Further embodiments show that the sensor system comprises at least one acoustic sensor and / or one radio wave-based sensor. The acoustic sensor is, for example, a sonar. The radio wave-based sensor is, for example, a radar.

[0027] In exemplary embodiments, the data processing unit is configured to detect, based on a multitude of successive sensor signals, whether the event constitutes an obstacle and whether the obstacle is static or moving. Static obstacles include, for example, (potentially uncharted) buoys or anchored vessels. For underwater vehicles, an obstacle can also be (potentially uncharted) sunken ships or reefs. In particular, an object that is not shown on a nautical chart to which the vessel has access can be considered an obstacle.

[0028] Moving obstacles include ships. In the case of underwater vehicles, moving obstacles can also be, especially large, living creatures such as whales or sharks, or even other underwater vehicles. For example, once a moving obstacle has been detected, the data processing unit can perform a course prediction, depth prediction, and / or speed prediction of the moving obstacle and estimate the risk of a collision based on the course and speed predictions.

[0029] This means that, based on knowledge of the vessel's own speed and direction, and knowledge of the speed and direction of the moving obstacle, the data processing unit can, for example, estimate an intersection point of both routes and / or a minimum distance that the unmanned vessel maintains from the moving obstacle. In particular, this makes it possible to maintain a prescribed minimum distance from the moving obstacle at all times, which can be considered a threshold for the risk of a collision. The propulsion and control unit can be controlled by the data processing unit in such a way that, in particular, a predetermined minimum distance between the autonomous vessel and the moving obstacle is maintained at all times.

[0030] In further embodiments, the data processing unit can access and evaluate nautical chart information. This means that the information recorded on the chart, such as buoys, is known to the autonomous vessel. The data processing unit can compare a detected static obstacle with the chart information. This provides navigation support, particularly by comparing the current position relative to the surrounding recorded fixed (static) obstacles. Furthermore, by detecting obstacles not shown on the chart, it is possible to draw attention to obstacles that could not be considered during initial route planning and thus may potentially lie within the unmanned vessel's path.

[0031] In exemplary implementations, the data processing unit is configured to prescribe the evasive maneuver in such a way that generally applicable maritime regulations are observed. These generally applicable maritime regulations can be the "International Regulations for Preventing Collisions at Sea 1972" (COLREGs) or other, for example, locally applicable maritime regulations. This allows the autonomous vessel to navigate maritime traffic like a manned vessel. This is advantageous for obtaining certification of the autonomous vessel by the maritime authorities.

[0032] In exemplary embodiments, the autonomous vessel is equipped with a signaling device. This can be an audio signal (e.g., ship's whistle, bell, or horn) or a visual signal such as flags or lights. The signaling devices enable compliance with the general rules of navigation. The data processing unit can control the signaling device in such a way that it indicates the vessel's operating mode, in particular, an evasive maneuver. This is prescribed, at least for certain maneuvers, by the general rules of navigation.

[0033] In further embodiments, the data processing unit determines a plurality of alternative changes of direction to execute the evasive maneuver and, based on an optimization criterion, selects the most advantageous maneuver from the plurality of possible evasive maneuvers to be executed. This allows for the comparison of different possible routes for the autonomous watercraft. Optimization criteria for evaluating the comparison can include, for example, travel time, route distance, the number of possible further evasive maneuvers, or any combination thereof. Alternatively, to save computing power, the first possible evasive route can be used based on a trial-and-error principle. Artificial intelligence methods can also be advantageously employed to determine the behavior.

[0034] Further embodiments demonstrate that the data processing unit can compare an alternative course for the evasive maneuver with other system variables of the autonomous watercraft. These system variables include, for example, the available energy storage for propelling the autonomous watercraft. Furthermore, the maneuverability or maximum speed of the autonomous watercraft can be considered a system variable. Thus, a route that is too long can be discarded, for example, due to insufficient energy reserves. Similarly, a route can be discarded if a course change is necessary that the autonomous watercraft cannot execute due to insufficient maneuverability. The same applies to an evasive maneuver where the speed cannot be increased sufficiently to prevent a collision or mitigate the impact of the event on the watercraft.can be reduced sufficiently.

[0035] In exemplary embodiments, the data processing unit is further configured to specify the pre-planned route in the absence of continuous or regularly recurring lane markings from the drive and control unit. Regular lane markings are defined as markings at regular intervals, with a maximum length of 10 times the autonomous watercraft. Thus, the navigation of an autonomous watercraft differs fundamentally from the navigation of an autonomous vehicle on the road. In the latter case, continuous or at least regularly recurring lane markings are present, which the sensors of the autonomous road vehicle can use for orientation.

[0036] Further embodiments show that the autonomous watercraft is an autonomous underwater vehicle and that the data processing unit can control the propulsion and control unit in such a way that the autonomous underwater vehicle follows a three-dimensional route in space. That is, unlike an autonomous surface vehicle, which follows a two-dimensional route in a plane, the data processing unit for an autonomous underwater vehicle should determine a three-dimensional route that also includes the diving depth.

[0037] Similarly, a method for operating an autonomous watercraft with a propulsion and control unit for maneuvering the autonomous watercraft is disclosed, comprising the following steps: a) detecting environmental information of the autonomous watercraft by means of a sensor system comprising one or more sensors; b) outputting at least one sensor signal corresponding to the environmental information for each sensor; c) controlling the propulsion and control unit such that the watercraft travels along a pre-planned route; d) detecting an event based on the sensor signal(s) and determining the risk of the event to the watercraft; e) controlling the propulsion and control unit such that the watercraft performs an evasive maneuver to avoid or mitigate the event, provided the risk exceeds a threshold;f) Training a machine learning model to control the autonomous watercraft using training data based on sensor signals and control signals to control the propulsion and control unit.

[0038] Also disclosed is a corresponding computer program comprising commands which, when the program is executed by a computer of an autonomous watercraft with a propulsion and control unit for maneuvering the autonomous watercraft, cause the computer to execute the procedure.

[0039] Preferred embodiments of the present invention are explained below with reference to the accompanying drawings. These show:

[0040] Fig. 1: a schematic block diagram of an autonomous watercraft with a data processing unit; and

[0041] Fig. 2: a schematic block diagram illustrating some of the data processing steps of the data processing facility.

[0042] Before exemplary embodiments of the present invention are explained in detail below with reference to the drawings, it should be noted that identical, functionally equivalent or equivalent elements, objects and / or structures in the different figures are provided with the same reference numerals, so that the description of these elements shown in different exemplary embodiments is interchangeable or can be applied to one another.

[0043] Fig. 1 shows a schematic block diagram of an autonomous watercraft 20 with a data processing unit 22. The autonomous watercraft further comprises a sensor system 24, which is connected to the autonomous watercraft by means of lines 26 in order to transmit the sensor signal(s) to the data processing unit 22. The sensor system comprises at least one sensor 24', optionally at least one further sensor 24", but preferably a plurality of sensors. The autonomous watercraft further comprises a drive and control unit 28. The data processing unit 22 is connected to the drive and control unit by means of further lines 30 for controlling the autonomous watercraft 20. Optionally, the sensor system 24 also has a direct line 32 for controlling the autonomous watercraft 20 to the sensor system 24.Optionally, the autonomous watercraft 20 has a signaling device 34 which is connected to the data processing unit 22 by means of a further line 36.

[0044] Fig. 2 shows a schematic block diagram of a section of the data processing system with the data processing unit 22. The data processing unit 22 receives sensor data 26 from a sensor system 24. The sensor data 26 are processed in a deterministic control module 40. The deterministic control module 40 is configured to output control data 42 so that the autonomous watercraft follows a predefined route. Furthermore, the deterministic control module 40 can detect an unforeseen event and adjust the control data 42 such that the autonomous watercraft performs an evasive maneuver.

[0045] Furthermore, the data processing unit 22 includes a model 44, optionally a plurality of models 44, 44', 44". The model 44(s) receive the sensor data 26 and the control data 42. The data processing unit 22 trains the model(s) 44, 44', 44" to control the autonomous watercraft using machine learning. The training data for the machine learning is based on the sensor signals 26 (input data) and the corresponding control signals 42 (output data). That is, the model(s) are trained using the sensor signals to output a result 46, 46', 46" for controlling the drive and control unit that reflects the control signals 42 as accurately as possible or even delivers a better result.

[0046] An optional decision module 48 can now determine and evaluate the quality of the results 46, 46', 46". This evaluation is optionally performed with regard to the quality of the control signals 42. If the quality of the results 46, 46', 46" is sufficient, the decision module 48 can use one or more of the results 46, 46', 46" to control the drive and control unit 28 via line 30. The decision as to which result(s) 46, 46', 46" are used to control the drive and control unit 28 can be made based on a decision criterion. For example, the result of the best model can be used, or a majority decision of the models can be made.

[0047] The decision module 48 can optionally have access to database 50 containing general navigation rules. The decision module 48 can compare the results 46, 46', 46" of models 44, 44', 44" with database 50 to ensure that the autonomous vessel operates in accordance with the general navigation rules. If the comparison reveals a violation of the general navigation rules, the decision module 48 can control the propulsion and steering unit 28 with the control signals 42.

[0048] Optionally, the data processing unit 22 includes an interface 52 for receiving external commands 54. The decision module 48 can replace the control signals 42 with the external commands 54 in order to control the drive and control unit 28 with the external commands 54. To train the models 44, 44', 44"", the control signals 42 are replaced by the external commands 54. If an emergency situation is detected by the data processing unit 22, the external command 54 can be overwritten using the current control signal 42. The control signal 42 is then used to train the models 44, 44', 44"".

[0049] In general terms, it is optionally possible for the decision module 48 to decide which training data, in particular input data, are used to train the model or models 44, 44', 44".

[0050] Although some aspects have been described in connection with a device, it is understood that these aspects also constitute a description of the corresponding process, so that a block or component of a device is also to be understood as a corresponding process step or as a feature of a process step. Similarly, aspects described in connection with or as a process step also constitute a description of a corresponding block, detail, or feature of a corresponding device.

[0051] The embodiments described above merely illustrate the principles of the present invention. It is understood that modifications and variations of the arrangements and details described herein will be obvious to other people skilled in the art. Therefore, it is intended that the invention be limited only by the scope of protection set forth in the following claims and not by the specific details presented herein by way of description and explanation of the embodiments.

[0052] Reference symbol list:

[0053] 20 autonomous watercraft

[0054] 22 Data processing facility

[0055] 24 sensor system

[0056] 26 (electrical) line / sensor signals

[0057] 28 Drive and control unit

[0058] 30 (electrical) line / control signals

[0059] 32 (electrical) line

[0060] 34 Signaling device

[0061] 36 (electrical) line

[0062] 40 deterministic control module

[0063] 42 control signals

[0064] 44 models

[0065] 46 results

[0066] 48 Decision module

[0067] 50 Database of maritime rules

[0068] 52 Interface

[0069] 54 external commands

Claims

Patent claims 1. Autonomous watercraft (20) with the following features: - a propulsion and control unit (28) designed to maneuver the autonomous watercraft (20), - a sensor system (24) comprising a sensor (24') or a plurality of sensors (24', 24"), wherein the sensor system (24) is configured to detect environmental information of the autonomous watercraft (20) and to output a corresponding sensor signal (26) for each sensor (24'); - a data processing unit (22) which is configured to control the drive and control unit (28) by means of control signals (42) such that the watercraft (20) travels along a pre-planned route; - wherein the data processing unit (22) is configured to detect an event based on the sensor signal or sensor signals (26) and to determine a risk of the event to the watercraft (20) and, if the risk exceeds a threshold, to control the propulsion and control unit (28) with the control signals (42) such that the watercraft (20) performs an evasive maneuver to avoid or mitigate the event; - wherein the data processing unit (22) is additionally equipped to train a model (44, 44' , 44“) for controlling the autonomous watercraft using training data based on the sensor signals (26) and on control signals (42) for controlling the drive and control unit (28).

2. Autonomous watercraft (20) according to claim 1, wherein the sensor signals (26) represent the input data of the model (44, 44', 44") and wherein the control signals (42) represent the results of the model (44, 44', 44").

3. Autonomous watercraft (20) according to one of the preceding claims, wherein the data processing unit is configured, after the model (44, 44', 44") has been trained with a plurality of training data, to determine a quality of the model (44, 44', 44") before further training using current sensor signals (26).

4. Autonomous watercraft (20) according to claim 3, wherein the data processing unit (22) is configured to control the propulsion and control unit (28) based on the results (46, 46', 46") of the model (44, 44', 44") when the quality of the model (44, 44") exceeds a threshold.

5. Autonomous watercraft (20) according to claim 4, wherein the data processing unit (22) comprises a database (50) containing general rules of navigation, wherein the data processing unit (22) is configured to compare the result (46, 46', 46") of the model (44, 44', 44") with the database to ensure that the autonomous watercraft (20) operates in accordance with the general rules of navigation and, in particular, if the comparison reveals a violation of the general rules of navigation, to control the propulsion and steering unit (28) with the control signals (42).

6. Autonomous watercraft (20) according to any one of the preceding claims, - wherein the data processing unit (22) includes an interface (52) for receiving external commands (54); - wherein the data processing unit (22) is configured to replace the control signals (42) generated by following the planned route or evasive maneuver with the external commands (54) in order to control the drive and control unit (28) based on the external commands (54); - wherein the data processing unit (22) is configured to replace the control signals (42) with the external commands (54) for training the model.

7. Autonomous watercraft (20) according to claim 6, - wherein the data processing unit (22) is designed to override the external command (54) in an emergency situation using the current control signal (42); - wherein the data processing unit (22) is configured to use the control signal (42) to train the model (44, 44', 44").

8. Autonomous watercraft (20) according to one of claims 6 or 7, wherein the data processing unit (22) is configured to receive information at the interface (52) to interrupt the training of the model.

9. Autonomous watercraft (20) according to one of the preceding claims, - wherein the data processing unit (22) is configured to train a plurality of different models (44, 44', 44") simultaneously.

10. Autonomous watercraft (20) according to claim 9, wherein the data processing unit (22) is configured to control the propulsion and control unit (28) with selected results (46, 46', 46") depending on the quality of any number of models (44, 44', 44") of the plurality of different models or after a predefined training period, depending on a decision criterion based on the results of the models (44, 44', 44") of the plurality of models.

11. Autonomous watercraft (20) according to claim 10, wherein the data processing unit (22) is configured to select the best model (44, 44', 44") from the plurality of different models as a decision criterion and to control the propulsion and control unit (28) based on the results of the best model (44, 44', 44").

12. Autonomous watercraft (20) according to claim 10, wherein the data processing unit (22) is configured to determine a majority decision as a decision criterion and to control the propulsion and control unit (28) according to a selection of the results based on that of the majority decision.

13. Method for operating an autonomous watercraft (20) with a propulsion and control unit (28) for maneuvering the autonomous watercraft (20) comprising the following steps: a) Detecting environmental information of the autonomous watercraft by means of a sensor system (24) comprising one or more sensors; b) Outputting a sensor signal (26) corresponding to the environmental information for each sensor (24'); c) Controlling the propulsion and control unit (28) by means of control signals such that the watercraft (20) travels along a pre-planned route; d) Detecting an event based on the sensor signal(s) and determining the risk of the event to the watercraft (20); e) Controlling the propulsion and control unit (28) with the control signals such that the watercraft performs an evasive maneuver to avoid or mitigate the event, provided the risk exceeds a threshold;f) Training a machine learning model to control the autonomous watercraft using training data based on sensor signals and control signals to control the propulsion and control unit.

14. Computer program comprising instructions which, when the program is executed by a computer of an autonomous watercraft with a propulsion and control unit for maneuvering the autonomous watercraft, cause it to execute the method according to claim 13.

Images