METHOD AND SYSTEM FOR UPDATING A CONTROL MODEL FOR AUTOMATIC CONTROL OF AT LEAST ONE MOBILE UNIT
Patent Information
- Application Number
- DE502018015325
- Authority / Receiving Office
- DE · DE
- Patent Type
- Patents
- Current Assignee / Owner
- Priority Date
- 2017-09-29
- Filing Date
- 2018-09-24
- Publication Date
- 2024-11-28
- Estimated Expiration
- 2038-09-24
AI Technical Summary
Existing methods for updating control models for autonomous vehicles are inefficient and costly, as they require extensive data collection in various driving situations, which can be time-consuming and resource-intensive.
A method and system where a central control unit generates a recording order for mobile units to collect specific sensor data, which is then used to generate an updated control model, allowing for targeted data collection and optimization through exploratory learning and cooperative acquisition, enabling the identification and optimization of weaknesses in automatic driving functions.
This approach enables faster and more efficient optimization of control models by prioritizing data collection based on identified weaknesses, reducing the need for extensive data gathering and improving the performance of autonomous driving functions.
Description
[0001] The present invention relates to a method for updating a control model for automatic control of at least one mobile unit. The invention further relates to a system for updating a control model for automatic control of at least one mobile unit.
[0002] Due to the diversity and complexity of the driving situations and environments encountered during the operation of an automatically controlled vehicle, the automated system cannot always perform the driving task with the same level of quality. Depending on the situation, for example, restrictions may occur in terms of comfort for the vehicle's occupants, or a route may not be completed in the optimal time.
[0003] US 2017 / 135621 A1 describes the collection of training data for autonomous driving and the training of a model based on the collected training data. EP 3 219 564 A1 describes a driver assistance system based on machine learning. US 2017 / 174221 A1 describes the analysis of driving characteristics using machine learning, and US 2010 / 063626 A1 describes the identification of a vehicle's environment using a knowledge database. US 2015 / 244826 A1 describes the prioritization of collected sensor data from an autonomous vehicle.
[0004] DE 10 2014 212 898 A1 proposes a method for the automated driving of an ego vehicle, in which an emergency trajectory for avoiding a collision is generated, among other things, on the basis of data collected by means of Car2Car communication.
[0005] EP 2 164 059 A1 describes a method for controlling a traffic situation in which sensor data is exchanged between several vehicles involved in the traffic situation. The vehicles' trajectories are then controlled based on this data.
[0006] There are also known methods in which test drivers use specially equipped vehicles to conduct targeted drives in problematic situations. The data collected is used to improve automated driving systems. However, this is associated with high costs and a significant time-consuming effort if the largest possible number of individual situations are to be recorded.
[0007] The present invention is based on the object of providing a method and a system of the type mentioned above, in which the automatic travel of a mobile unit can be optimized particularly efficiently and comprehensively.
[0008] According to the invention, this object is achieved by a method having the features of claim 1 and a system having the features of claim 10. Advantageous embodiments and further developments emerge from the dependent claims.
[0009] In the method according to the invention of the type mentioned above, a central control unit generates a data acquisition task and transmits it to the mobile unit. The mobile unit comprises sensors, and the data acquisition task comprises conditions for the acquisition of sensor data sets using the sensors. The mobile unit acquires the sensor data sets using the sensor data based on the data acquisition task, generates transmission data based on the acquired sensor data sets, and transmits the transmission data to the central control unit.The central control unit receives the transmission data and, based on the received transmission data, generates an updated control model for the automatic control of the mobile unit, wherein in an exploratory learning phase a plurality of candidate trajectories for the automatic control of the mobile unit are generated, and a trajectory determined to be safe and not preferred is driven, wherein the non-preferred trajectory is re-evaluated based on the sensor data sets acquired during execution, the new evaluation results generated in this way are transmitted to the central control unit and the control model is optimized for automatic trip planning.
[0010] This advantageously allows for a particularly targeted collection of data and an optimization of the control model can be carried out more quickly and efficiently than with the known methods.
[0011] The process is carried out in several phases. In a first phase, a need for optimization can be identified. In a second phase, a cooperative data acquisition and learning task can be determined and transmitted to participating mobile units. In a subsequent phase, the mobile units, or rather based on data collected by these units, conduct cooperative exploration, measurement, and learning. In a further phase, the results can be transmitted to the mobile units and an optimized functionality can be activated.
[0012] A control model within the meaning of the invention refers to a unit of a control system used for the automatic control of a mobile unit. The control model serves to assess information about the current driving situation of the mobile unit, in particular by evaluating sensor data, and to make a decision about a reaction of the mobile unit or to determine parameters for an automatic driving function. In particular, it is provided that the automatic control is carried out by means of modules that control specific devices of the mobile unit, with each module being assigned a control model. For example, a control model can be provided for use with a neural network or a machine learning method.
[0013] The mobile unit is, in particular, a motor vehicle. The method can involve a plurality of mobile units, in particular motor vehicles, which can be organized, for example, as a connected fleet of vehicles.
[0014] In particular, the central control unit is comprised of an external server. The central control unit may also be another vehicle-external device and, for example, be comprised of a distributed computer system. The central control unit may have an interface via which data can be transmitted to the mobile unit or the plurality of mobile units.
[0015] The transmission of data between the mobile unit and the central control unit can take place in a manner known per se by means of a data connection between the mobile unit and the central control unit, in particular wirelessly. For example, the connection can be established via a local area network or a larger network, such as the Internet. Furthermore, the connection can be established via a telecommunications network, such as a cellular network, or a wireless local area network (such as WLAN). Furthermore, the data connection can be established by connecting a data cable. The connection can also be established via another unit that can itself establish a connection to the central control unit.For example, a data connection may exist between the mobile unit and a mobile phone connected to the Internet, such as a data cable or a wireless connection, such as Bluetooth. In particular, the connection may be established via the Internet.
[0016] In the method, a recording order is transmitted from the central control unit to the mobile unit. This order includes conditions for recording data using the mobile unit's sensors. These conditions define, for example, in which driving situations which types of sensor data should be recorded. This means that the recorded sensor data sets can be created based on the conditions of the recording order. This does not necessarily mean that the mobile unit's sensors only record sensor data when the conditions of the recording order are met. Rather, continuous recording can also be provided, with the data to be recorded being selected based on the recording order, in particular based on the conditions it encompasses.For example, the acquired sensor data can be stored in a ring buffer, which can then be analyzed based on the acquisition task and the conditions to generate the relevant sensor data sets.
[0017] The transmission data generated from the sensor data sets can be formed in various ways and, for example, can include the data from the sensor data sets. The sensor data sets can be further processed, for example, to anonymize personal data or to perform a procedural step for generating a provisionally updated control model at this point, for example, to relieve the central control unit or to enable it to further process the transmission data provided by the mobile unit.
[0018] The transmission data can be transmitted to the central control unit in a manner known per se. In particular, it can be provided that a data connection is established between the mobile unit and the central control unit at specific times, for example, after the completion of a journey, in particular to transmit transmission data and / or a recording order. Furthermore, a data connection established at periodic intervals or a continuous one can be provided.
[0019] The updated control model is also generated in a manner known per se from the transmission data, in particular by means of a machine learning method. It is based, in particular, on an earlier version of the control model, for example, the one used by a module of the mobile unit or another control model. Furthermore, the updated control model can be based on other models or be completely recreated from the transmission data.
[0020] The method according to the invention can be implemented in such a way that cooperative learning is possible to identify problematic situations in automated driving and, in particular, to generate strategies and data cooperatively with other vehicles in order to optimize their driving functions based on this knowledge. By using a central control unit, many vehicles in a fleet can be used for data collection. The potentially very large amounts of data generated in this process can be collected more specifically and their overall volume reduced by specifically requesting specific sensor data in specific situations.
[0021] The slightly different conditions encountered during each journey by a vehicle in a fleet, for example, depending on the current weather, time of day, and the exact location of the vehicle, allow for comprehensive data collection under different conditions. Data-driven autonomous driving functions can thus be optimized particularly comprehensively.
[0022] In one embodiment of the method according to the invention, a requirement report is received by the central control unit, and the data entry order is generated based on the requirement report. This advantageously allows for the generation of the updated control model to respond to identified weaknesses in the existing system.
[0023] In the invention, the requirement report can, for example, comprise information about a functioning of the control model, in particular an evaluation of the function of an automatic driving function which is controlled, for example, by an automatic control module using a control model. In this way, it can be detected, for example, whether the automatic control is not functioning optimally. The requirement report further comprises information about a context in which the non-optimal functioning occurs. In particular, it is determined in which situations and / or at which locations a weakness occurs. An identified weakness can, for example, relate to a location, a class of locations, a situation, or a class of situations. The method according to the invention now makes it possible to deal with the weaknesses thus identified and to use the available resources as efficiently as possible.
[0024] Non-optimal functioning may be caused by various factors, whereby in particular an analysis step may be provided in which it is determined which control model for controlling the affected automatic driving function is responsible for non-optimal functioning.
[0025] The information contained in the demand report about the context in which the suboptimal driving behavior occurred can, for example, relate to a position, a traffic situation, or a state of the mobile unit. This can include information about the traffic infrastructure being used, such as the type of road, road layout, intersections, junctions, traffic lights, and priority rules; about a driving maneuver performed, such as an overtaking maneuver or a turn; about an operating state of the mobile unit, such as operating temperature, speed, or acceleration; or about climatic conditions, such as precipitation, fog, night, twilight, or strong sunlight.
[0026] A weakness or suboptimal functioning of an automated driving function can be identified based on one of the following factors, for example: the occurrence of a critical or unsafe system state, dangerous or overly dynamic maneuvers, interventions by a safety function in the automatic control system, uncomfortable driving behavior, unsafe (virtual) sensors, or a detected misperception. Furthermore, suboptimal coordination of individual components along the automated driving processing chain can be identified, particularly in one of the elements: sensors, perception, recognition, situational awareness, production, planning, and actuators. A weakness can also be identified if an intervention by the driver is detected.
[0027] The demand report is generated primarily using sensor data collected during an automated journey by the mobile unit. This sensor data can be collected, for example, by sensors on the mobile unit itself or by sensors external to the vehicle. The evaluation of this sensor data and the generation of the demand report can also be carried out by different units.
[0028] In particular, to identify a weakness in the automatic driving functions, a parameter measured by sensors is compared with a threshold value. If the value exceeds or falls below the threshold value, this can be interpreted as an indication of the presence of a weakness. For example, the parameter can be an acceleration in the longitudinal and / or transverse direction relative to the vehicle axis; in this case, the identification of non-optimal functioning can be carried out based on the occurrence of jerky acceleration or deceleration. In another example, the parameter can be a speed of the mobile unit, which is compared with the speeds of other mobile units on the same route.For example, if it is determined that the automatically controlled mobile unit covers the same distance significantly slower than other mobile units, this can be concluded that the system is not functioning optimally.
[0029] The identified weaknesses of the automatic driving functions can in particular be so-called " Corner Cases" that can occur in extreme cases, especially when certain parameters assume extreme values at the same time. Furthermore, the weaknesses can represent situations in which an improvement of the control model for automatic control is necessary.
[0030] In a further embodiment, the demand notification is generated by the mobile unit, using, in particular, driving data acquired by the mobile unit's sensors. This advantageously allows the central control unit to receive information directly from the mobile unit regarding the need to update the control model.
[0031] In a further development, the demand notification is generated by a traffic monitoring device. This is particularly automatic. This allows additional external vehicle devices to be used to identify weaknesses in the automatic control system.
[0032] The traffic monitoring device is, in particular, a device external to the vehicle and fixed relative to the road used to collect data on traffic in a specific spatial area. Communication between the mobile unit and the traffic monitoring device can also be provided (e.g., via " Car2X"), whereby, for example, an automatically controlled vehicle can be identified. Traffic monitoring devices can be provided, for example, at intersections, construction sites in the vicinity of schools and public facilities, at toll booths or along public roads. For example, such a device can record the movement of the mobile unit, particularly in connection with other mobile units, and an evaluation can be carried out. For example, it can be detected whether the mobile unit is following an optimal trajectory in order to cover a route as safely and quickly as possible. For example, a traffic monitoring device can register that an automatically controlled vehicle tends to drive in a zigzag pattern in certain situations.
[0033] In further examples, a weakness can be identified by another unit, such as a subunit of the central control unit or an external server. Data collected by the mobile unit and / or traffic monitoring devices can be evaluated to determine the occurrence of a weakness and information about the context in which this weakness occurs. In this case, the demand notification can be generated by the other unit and transmitted to the central control unit.
[0034] The identification of optimization needs for an automatic function or control model can be achieved in different ways. First, the occurrence of the weakness itself can be detected based on the factors described above, such as sudden changes in speed or automatic interventions of a safety function. Furthermore, " Data mining" can be used to determine in which situations the weaknesses occur. Driving data can be examined to uncover correlations between system weaknesses and environmental parameters, such as geographical parameters, climatic conditions, traffic, road type, speed, time of day and year, etc. Furthermore, traffic observation can be used to detect anomalies, such as traffic jams or heavy traffic. In addition, an assessment of the reliability of decisions made by individual modules along the processing chain can be carried out at runtime. Plausibility methods can be applied for individual components or the entire system. Furthermore, the frequency with which switching between redundant system components takes place can be determined. Physical measured variables can also be evaluated, such as longitudinal and lateral accelerations, particularly relative to the surrounding traffic or the road layout.Occupant feedback can be measured, such as through driver observation, manual intervention detection, and / or explicit feedback. Furthermore, a parameter can be measured and compared with a prediction for the same parameter from a previous time step to assess the reliability of the prediction.
[0035] During training, the central control unit receives a large number of requirement reports. The requirement reports are weighted, and the data entry order is generated based on the weighted requirement report. This advantageously allows for the management of multiple requirement reports and the coordination of the data entry order.
[0036] For this purpose, reported weaknesses, particularly based on requests for improvement, are collected and evaluated by the central control unit. Aspects of a control model that require urgent optimization can be identified, for example, based on the frequency with which a functional weakness occurs, on a strong correlation between a functional weakness and certain environmental parameters, or on the severity of identified weaknesses. For this purpose, the severity of a weakness can be assessed using a defined scale, and more serious weaknesses can be addressed with priority. The weighting of requests for improvement can therefore be based on various assessment criteria, and combinations of criteria are also possible.
[0037] This process of collecting and evaluating data is carried out fully automatically, continuously, and parameterized by the central control unit. In particular, methods from the field of " Big Data ", where large amounts of data of different types are processed at high speed.
[0038] After identifying a functionality to be optimized or a control model to be updated, the central control unit generates a data acquisition task and transmits it to the mobile unit. The data acquisition task is, in particular, an explicit learning task that is transmitted, for example, to the vehicles in a fleet. The conditions for acquiring sensor data sets included in the data acquisition task are specifically designed for the respective application.
[0039] The acquisition task can refer to a module level, where the module serves to control a specific automatic driving function within the mobile unit's automatic control system. The control model is used by the module to determine target variables for the automatic control system and / or to make decisions for the automatic control system. Various non-optimal behaviors can occur within a module, such as high uncertainty, misclassifications, or erroneous predictions. For optimization purposes, specific data can be collected for the situation in which the weakness of the control model arises. This data can include, for example, image data, map data, and / or various sensor data.A local update of the control model may be performed, in which case the transmission data may include data from this local update, and the central control unit may perform a consolidation in which the updates performed locally (for example, in several different mobile units) are combined to generate an updated control model.
[0040] Alternatively or additionally, the data acquisition task can be related to a planning level of the automatic control system. In this case, the weakness concerns a control model that controls an automatic driving function based on the outputs of several modules. A control model at the planning level is therefore located at a higher level than the module level.
[0041] An automatic route planning function can be configured to select from a range of different candidates for trajectories and / or decisions for automatic control in relevant situations. The best candidate is selected. To optimize such a function, an exploration phase is provided, during which the option identified as the "best candidate" is not chosen, but rather another trajectory that is safely navigable but determined as non-preferred. This means that other candidates besides the actually preferred option are tried out and can be re-evaluated based on the data collected during execution. The new evaluation results generated in this way are transmitted to the central control unit, and the control model is optimized at the automatic route planning level based on the newly generated knowledge.
[0042] Furthermore, optimization can be performed at a system and redundancy level, whereby various system settings are exploratively optimized with respect to a specific situational weakness by systematically collecting decision parameters in the internal estimation of the individual module results along the processing chain and at the redundancy level. In such exploratory optimization, similar to the one described above, decisions can be made in automatic control for non-preferred options, as long as these are assessed as sufficiently reliable.
[0043] It may be intended that the data collection order is transmitted to all vehicles in a fleet. This can be done, in particular, for the purpose of collecting data and for training and optimizing a control model in a " Shadow Mode", in which the control model or a sub-functionality thereof neither actively intervenes in the driving system nor changes the current configuration of the automatic driving function. This ensures that no uncontrolled changes to the automatic control occur.
[0044] Furthermore, the data collection order may concern a portion of a vehicle fleet, particularly vehicles that are technically equipped in a special way or whose drivers have consented to secure, exploratory data collection. In this case, data on the behavior of the automatic control system can be collected and modified for the purposes of data and knowledge creation and in the area of safe driving behavior. In particular, such a portion of the vehicle fleet may include special research and development vehicles driven by test drivers and equipped with special development hardware and / or special sensors.
[0045] As the process continues, the participating mobile units conduct cooperative exploration and measurement, and cooperative learning can also occur. This is coordinated by the central control unit.
[0046] The acquisition task transmitted to the mobile units is executed. According to the method, sensor data sets are acquired and transmission data is generated, which is then ultimately transmitted to the central control unit. In particular, the conditions for the acquisition of sensor data sets included in the acquisition task are taken into account. In particular, the participating mobile units can register with the central control unit, whereby, in particular, a data connection is established and the acquisition task is transmitted.
[0047] In a further development, the conditions of the acquisition order for the acquisition of sensor data sets include data on traffic situations. The sensor data sets can thus advantageously be acquired specifically for specific situations in which a weakness of the existing system has been identified. The data on traffic situations relate, for example, to features of a traffic infrastructure such as intersections, road layout, or road types. They can also relate to specific driving maneuvers and the status of the mobile unit, for example, based on vehicle data. They can also relate to other mobile units that are located in the vicinity of the mobile unit.
[0048] The acquisition order can alternatively or additionally include further conditions, for example, information about the relevant functionalities, parameters for defining the situations for which sensor data sets are to be acquired (e.g., triggering events, defining framework conditions for the learning process to be carried out), system parameters to be optimized, triggering conditions, and the type and quality of the data to be acquired. Furthermore, the acquisition order can include information about a specific type of acquisition of sensor data sets and their processing for generating transmission data, such as a specific algorithm, defining parameters, or learning rates for a specific training method. Furthermore, an update frequency can be defined, which defines the transmission of the transmission data to the central control unit. Completion conditions can also be defined (" Definition of Done"), where a final configuration is defined that defines a final transmission of the transmission data to the central control unit. During the acquisition of the sensor data sets and the generation of the transmission data, metainformation can also be collected, the type of which is determined by the conditions of the acquisition job.
[0049] The information collected and / or generated by the mobile unit can be structured using a functional variant evaluation matrix to provide the data in a systematic and easy-to-process manner.
[0050] During the automatic control of the mobile unit, sensor data is collected by its respective sensors and situations or situation classes are recognized for which the recording task provides for the recording and / or transmission of data or the triggering of a learning process.
[0051] In particular, the data collected by the mobile unit's sensors is stored in a ring buffer that, for example, retains the data from the last 10 minutes at any given time. The retained data can include, for example, sensor data, map data, and / or data available via a CAN bus. If a relevant situation or situation class is detected, data from the recent past can also be retrieved from the ring buffer.
[0052] Furthermore, it can be provided that if several mobile units are involved, the recorded data is synchronized.
[0053] The acquisition order specifically defines conditions for the mobile unit to complete the learning phase, in particular, terminating the acquisition of sensor data sets and / or transmitting the transmission data. For example, this can occur at a specific, predetermined time. Furthermore, a termination condition can be provided, or a signal can be transmitted from the central control unit to the mobile unit to trigger the completion of the learning phase.
[0054] When the completion criterion for a data acquisition task is reached, the phase of acquiring sensor data sets and generating transmission data for all participating mobile units is terminated. The information acquired, which was transmitted to the central control unit via the transmission data, is aggregated and processed by the central control unit. The updated control model is generated, with particular emphasis on integrating the new information into the respective function in order to make the respective function more robust and / or optimize it. Furthermore, the updated control model can be compressed. Newly acquired knowledge can be identified and stored. Existing, older data can be used in addition to the data obtained through the data acquisition task to generate the updated control model.
[0055] During further development, the central control unit uses the transmission data to train a previous control model to generate the updated control model. In particular, a machine learning method is used. This advantageously allows for targeted optimization of the control model.
[0056] During training, an integrated control model is trained to generate the transmission data by the mobile unit. In particular, the transmission data includes data from the integrated control model trained in this way. Here, too, a machine learning method is used. This advantageously allows preprocessing of the transmission data in such a way that, for example, anonymized data transfer can be ensured. Furthermore, a cooperative learning process can be designed particularly efficiently in this way. In particular, the transmission data in this case includes data from a control model that has already been trained using a training method, thus providing a preliminarily trained control model.
[0057] This means that the training process can be carried out by the mobile unit itself and / or by the central control unit. This allows for a particularly flexible system design, with different training steps being carried out by the most suitable unit in each case.
[0058] The central control unit performs consolidation based on the transmitted data, whereby transmitted data from training processes in the mobile unit can also be used. In particular, the results of training in multiple mobile units can be consolidated, for example, by averaging certain parameters or other known distributed learning methods.
[0059] Further steps may be performed that are required for applying the updated control model to the mobile unit. For example, a security evaluation may be performed, and the updated control model may be optimized for a specific target hardware, such as by reducing its size or removing unused subfunctionalities. Furthermore, structural optimization may be performed to achieve optimal hardware utilization.
[0060] In a further development, the updated control model is transmitted to the mobile unit, and automatic control of the mobile unit is performed based on the updated control model. This advantageously benefits the mobile unit from the control model optimized based on the sensor data sets, enabling improved control.
[0061] The updated control model can first be transferred to an approval process, which ensures that only approved updates or changes are made to the control model of the mobile unit. Subsequently, a "Rollout ", in which the updated control model is transferred to the mobile unit and can then be used for automatic control.
[0062] In further embodiments of the invention, devices of an off-board infrastructure can be used alternatively or in addition to the use of mobile units for acquiring sensor data sets and generating transmission data. As described above, a demand notification can be generated by a traffic monitoring device. However, additional data can also be acquired, for example, by a device for observing traffic along a specific route. In this way, cooperative learning based on as many information sources as possible can be realized.
[0063] For example, driving functions can be used by devices of a traffic monitoring infrastructure, such as cameras and / or communication devices on a road or intersection, to capture sensor data sets for the method according to the invention and to provide suitable transmission data. For this purpose, these devices can be optimized, for example, to perform certain preprocessing steps based on captured data.
[0064] Facilities within such an infrastructure can, for example, identify a need for optimization and generate a corresponding request and transmit it to the central control unit. Subsequently, a recording order for similar situations can be generated and transmitted to mobile units. The conditions can be configured in such a way that, in addition to the specific situation monitored by the infrastructure facility, such as a specific intersection, sensor data sets for comparable situations are also recorded, for example, at similar intersections not monitored by a fixed infrastructure.
[0065] Furthermore, a need for optimization can be identified in a situation without infrastructure monitoring, and a data acquisition task can be created for a situation monitored by an infrastructure facility. The knowledge gained can then be transferred to semantically similar situations, such as comparable intersections. The updated control model can be created in such a way that optimization using data from monitored situations can also be performed for a non-monitored situation.
[0066] The system according to the invention for updating a control model for an automatic control of at least one mobile unit comprises a central control unit which is designed to generate a recording order and to transmit it to the mobile unit; wherein the mobile unit comprises sensors and the recording order comprises conditions for the recording of sensor data sets by means of the sensors; wherein the conditions define in which driving situations which types of sensor data are to be recorded, the sensors of the mobile unit are designed to record the sensor data sets based on the recording order, to generate transmission data based on the recorded sensor data sets, and to transmit the transmission data to the central control unit;wherein the central control unit is configured to receive the transmission data and, based on the received transmission data, to generate an updated control model for automatically controlling the mobile unit. The system is further configured to generate a plurality of candidate trajectories for automatic control, and to travel a trajectory determined to be safe and non-preferred, wherein the non-preferred trajectory is re-evaluated based on the sensor data sets acquired during execution, the new evaluation results thus generated are transmitted to the central control unit, and the control model is optimized for automatic journey planning.
[0067] The system according to the invention is particularly designed to implement the above-described method according to the invention. The system thus has the same advantages as the method according to the invention.
[0068] In particular, the system comprises a plurality of mobile units, in particular a plurality of motor vehicles.
[0069] The invention will now be explained using embodiments with reference to the drawings. Figures 1A and 1B show embodiments of the system according to the invention, Figure 2 shows an embodiment of the method according to the invention and Figure 3 shows an example of exploratory recording of sensor data of an automatic journey.
[0070] With reference to the Figures 1A and 1B Embodiments of the system according to the invention are explained.
[0071] A mobile unit 1, in this case a vehicle 1, comprises a computing unit 4 coupled to a detection unit 3, actuators 9a, 9b, 9c, and a vehicle interface 5. The detection unit 3 comprises sensors 3a, 3b, 3c, 3d. In the example, the actuators 9a, 9b, 9c comprise devices for an engine control, a steering system, and brakes of the vehicle 1. The sensors 3a, 3b, 3c, 3d are formed in a manner known per se and, in the example, comprise a camera, a laser scanner, ultrasonic distance sensors, and a device for detecting user inputs.
[0072] Using the vehicle interface 5, a data connection (indicated by a dashed line) can be established, at least temporarily, to a server interface 6 of a central control unit 2, in this case an external vehicle server 2. The server 2 includes an update unit 7 coupled to the server interface 6.
[0073] In further embodiments, the vehicle 1 and the server 2 can be constructed differently. In particular, the equipment of the vehicle 1 with sensors 3a, 3b, 3c, 3d and actuators 9a, 9b, 9c can be designed differently.
[0074] As in Figure 1BAs shown, the system can comprise a plurality of mobile units 1, which can at least temporarily establish a data connection to the central control unit 2. In this exemplary embodiment, a traffic monitoring device 8 is also provided, for which a data connection to the central control unit can also be established. The traffic monitoring device 8 is not shown in detail here, but analogously to the mobile units 1, it comprises a detection unit with sensors 3a, 3b, 3c, 3d as well as a computing unit 4 and an interface for establishing the data connection. The number of mobile units 1 and traffic monitoring devices 8 can of course vary depending on the specific embodiment.
[0075] In further embodiments, it can be provided that data connections of the mobile units 1 to each other and / or to the traffic monitoring devices 8 can be established, for example in a manner known per se by means of a Car2Car or Car2X process.
[0076] The data connections can be permanent or established at specific times, for example, periodically or on request. They are established in a conventional manner, in particular through a wireless connection, such as a mobile network, a local computer network (e.g., Wi-Fi), or by means of another device through which the data connection can be established, such as a mobile phone. In particular, the connection is established via a computer network such as the Internet.
[0077] Furthermore, it is particularly provided that the establishment of the data connection is linked to an authentication process. A secure connection can be established between the server 2 and the vehicle 1 or the traffic monitoring device 8. This is done, for example, using a login process.
[0078] With reference to Figure 2 An embodiment of the method according to the invention is explained. A dashed line indicates the separation between method steps that are carried out in vehicle 1 (reference numeral 21) and those that are carried out in server 2 (reference numeral 22).
[0079] In a first step S1, an automatic drive is carried out with the vehicle 1. In this process, various automatic driving functions are controlled by modules that make driving decisions and / or determine parameters for controlling the drive using control models. Figure 1A In the embodiment of the system according to the invention shown, the modules with the control models are included in the computing unit 4 and control signals are generated and transmitted to the actuators 9a, 9b, 9c, by means of which the actual control of the vehicle, in particular its movement, is then carried out.
[0080] In a further step S2, a weakness is detected, i.e., a case of non-optimal functioning of the automatic control. In the example, this detection takes place using sensor data from vehicle 1, which, among other things, detects accelerations in the left and transverse directions relative to the longitudinal axis of the vehicle. The acceleration is compared with a threshold value, and if it is exceeded (i.e., for example, in the case of sudden braking or a sudden deviation from the direction of travel), this is interpreted as the occurrence of a weakness. Alternatively or additionally, further methods for detecting a weakness can be used, for example, a comparison with further sensor data, which can, for example, achieve a more precise detection of the traffic situation. In further exemplary embodiments, other methods can be provided to detect weaknesses in the automatic driving.
[0081] The detected weakness is assigned to a module for automatic control, in particular, the assignment is made to a specific control model of the module. For example, a weakness for a specific module can be detected with a control model for planning a curve, an overtaking maneuver, maintaining a certain distance from other road users, or for a lane assistant. Other examples include a module for controlling following a vehicle ahead, a module for executing a turning maneuver, a module for driving through a construction zone, and / or a module for executing a parking maneuver. The affected control model should then be trained in order to optimize it.
[0082] For this purpose, a demand report is generated in step S3 and transmitted to server 2. The demand report contains information about the detected weakness and the situation in which it was detected. For this purpose, the acquired sensor data is evaluated, and the method can provide different ways of generating the demand report. In particular, the information content transmitted with the demand report can be defined differently.
[0083] It is crucial that information is transmitted about the situations in which suboptimal automatic control occurs and how this affects automatic driving. Based on this information, the control model responsible for the weakness can be identified and it can be determined in which situations retraining the control model is necessary.
[0084] In further embodiments, the demand message can be generated by a traffic monitoring device 8 and transmitted to the server 2. This can be done in a manner analogous to the procedure described above. In further embodiments, a demand message can be transmitted between multiple vehicles 1 and / or traffic monitoring devices 8, whereby, for example, a coordination of multiple demand messages can take place, such as a confirmation of a suboptimal functioning of an automatic driving function by multiple devices and units.
[0085] In a step S4, the demand message is received from the server 2. In the exemplary embodiment, a plurality of demand messages are received, which were generated by one or more vehicles 1 and / or one or more traffic monitoring devices 8. The server 2 performs preprocessing, for example, determining how many demand messages were received for similar situations and / or the same control model. Alternatively or additionally, the severity of the demand messages can be assessed. Based on the number and / or severity of the demand messages, they can be weighted to determine which control models should be optimized through training and which data is required for this training.
[0086] In step S5, a data acquisition task is generated and transmitted to vehicle 1. The data acquisition task includes conditions under which sensor data should be acquired in order to perform the desired training of a control model. The conditions included in the data acquisition task particularly relate to parameters that are to be used to identify a situation for which the control model is to be trained.
[0087] In the exemplary embodiment, the acquisition request includes information about specific functionalities, in particular a control model to be trained. It also includes situation parameters that serve, for example, as triggers and defining framework conditions for the training process to be carried out, as well as system parameters to be optimized. It also includes information about how the sensor data is to be acquired, for example, with which spatial and / or temporal resolution, which quality, and in which type. Furthermore, information about a specific training method can be included, such as an algorithm, training parameters, or learning rates. Furthermore, information about which metadata is to be collected during the learning phase can be included.
[0088] In the exemplary embodiment, the acquisition job further comprises information about a frequency with which data is to be transmitted to the server 2, as well as a definition for the conditions which define a completion of the acquisition job and a final transmission of data to the server 2.
[0089] In step S6, vehicle 1 receives the acquisition request and acquires sensor data sets using acquisition unit 3, taking into account the conditions encompassed by the acquisition request. In particular, sensor data sets are continuously acquired and stored in a ring buffer that stores, for example, the sensor data sets from the last 10 minutes. If it is detected that the conditions of the acquisition request are met—that is, in particular, if vehicle 1 recognizes the presence of a specific driving situation for which a control model training is to be performed—corresponding data for the recent past can also be retrieved from the ring buffer.
[0090] Based on the acquired sensor data sets, transmission data is generated in a further step S7 and transmitted from vehicle 1 to server 2. The transmission data is generated depending on the acquisition request. In particular, the conditions encompassed by the acquisition request define criteria for defining a situation in which, for example, a weakness of an automatic driving function or a control model occurs. Function variants are recorded for the criteria thus specified, which include, for example, observations or variants learned from the sensor data sets. In this way, the transmission data can be generated such that suitable data can be transmitted for multiple acquisition requests, whereby overlapping criteria can be provided for different situations.For example, the situations "intersection with pedestrians at dusk" and "overtaking maneuver on a country road at dusk" may have common criteria ("at dusk") for which function variants should be transmitted to the server.
[0091] In the exemplary embodiment, the transmission data comprises an evaluation matrix which is formed in the following way: Criterion A ... Criterion Z Functional variant 1 a 1 ... Z 1 ... ... ... ... Functional variant n to ... Zn
[0092] "Criterion A" to "Criterion Z" denote the criteria that are specified in particular by the data acquisition task, "Functional variant 1" to "Functional variant n" denote the learned variants and "a 1 " to "an " denote the values recorded or learned for variants.
[0093] In further embodiments, the sensor data sets can alternatively or additionally be processed in a different way and provided via the transmission data. It is important that the transmission data allow a training or learning process for the control model for the automatic driving situations defined based on the acquisition task.
[0094] In further embodiments, it can be provided that the vehicle 1 already performs a learning process for the control model, and the transmission data includes information about the newly trained control model. For this purpose, a mode of the control model can be activated in the vehicle 1 in which it does not actively intervene in the control and can be modified by the learning process without thereby altering safety-relevant driving functions. Furthermore, a learning process of the vehicle can be permitted to a certain extent during active driving, with safety aspects being continuously monitored.
[0095] The time of generation and / or transmission of the transmission data can be determined in various ways. For example, this can always occur after a completed journey or when a data connection to the server can be established. Furthermore, the recording order can include provisions regarding when and under what circumstances the generation and / or transmission should take place.
[0096] In further embodiments, a transmission of data between multiple vehicles 1 is also provided. For example, sensor data sets and / or transmission data can be transmitted between the vehicles 1. This allows for more robust data transmission or allows learning processes to be performed at the level of the vehicles 1 using a larger database.
[0097] In a step S8, the server receives the transmission data, and in a step S9, an updated control model is generated. In the exemplary embodiment, this generation is performed using a machine learning method in which the information contained in the transmission data is used to determine the function of a control model in specific situations. In particular, the generation of the updated control model is preceded by a consolidation of the transmission data, possibly from a plurality of vehicles 1 and / or traffic monitoring devices 8, whereby distributed learning methods can be used.
[0098] The exemplary embodiment assumes a control model that is distributed to a plurality of vehicles 1 and is also available on the server 8. A learning method is used to update the control model, and the control model present on the server 8 is trained based on the knowledge learned by the vehicles 1.
[0099] In further embodiments, training can be performed in other ways. For example, training of the control models available in each of the vehicles 1 can be performed, and parameters of these control models can be transmitted to the server 8 via the transmission data. When the updated control model is generated by the server 8, a comparison can be made between the control model available on the server 8 and the control models trained in the vehicles 1; for example, an average of the parameter values can be performed.
[0100] In step S10, the updated control model is transmitted from server 2 to vehicle 1. In further embodiments, this step can be preceded by an approval process, in which, in particular, the safe functioning of an automatic driving function is ensured using the updated control model. This can be done, for example, by a central location; furthermore, automated procedures can be performed by server 2 for the approval of the updated control model. This is designed, in particular, in accordance with specific legal requirements.
[0101] In a step S11, vehicle 1 receives the updated control model and performs an automatic drive using the updated control model. In the exemplary embodiment, the updated control model is used instead of the old control model for the respective assigned module for an automatic driving function. In further exemplary embodiments, it can be provided that a backup copy of the old control model is retained, which can be reused in the event of a detrimental functioning of the updated control model.
[0102] With reference to Figure 3 An example of exploratory sensor data acquisition during an automated journey is explained. This particularly concerns a learning phase in which vehicle 1 acquires sensor data sets based on the acquisition task, which are then used to generate the transmission data.
[0103] Figure 3shows an automatically controlled first vehicle 10, which in the illustrated case has received a data acquisition order and is capturing corresponding sensor data sets. In the direction of travel (shown as arrow R) in front of the first vehicle 10 is another vehicle 11 that is to be overtaken.
[0104] During the automatic execution of the overtaking maneuver, a trajectory is determined along which the first vehicle 10 is to move in order to overtake the other vehicle 11. For this purpose, a plurality of potentially drivable trajectories 12a, 12b are generated in the exemplary embodiment, two of which are shown in the figure as examples and which are referred to below as candidates 12a, 12b. In a further step, the candidates 12a, 12b are evaluated, determining, among other things, how safely they can be driven, how efficiently the travel along the trajectories can be carried out, and how long the overtaking maneuver will require.
[0105] During an exploratory learning phase, the exemplary embodiment provides for using a different trajectory 12a, 12b rather than the trajectory 12a, 12b identified as preferred. This ensures that this other trajectory 12a, 12b can also be safely navigated; however, slightly different parameters can be tested in order to train the corresponding control model based on the sensor data sets acquired when navigating these trajectories 12a, 12b.
[0106] This method can be used in particular to carry out learning at the planning level, that is to say at a level for coordinating several modules for the automatic control of the journey.
[0107] In further embodiments, further methods, in particular from the field of machine learning, can be used to acquire suitable sensor data sets and / or generate transmission data based on the acquisition order, wherein, if necessary, a training of a control model can already be carried out in a mobile unit 1. List of reference symbols
[0108] 1Mobile unit; vehicle 2Central control unit; server 3Detection unit 3a, 3b, 3c, 3dSensors 4Computing unit 5Vehicle interface 6Server interface 7Update unit 8Traffic monitoring device 9a, 9b, 9cActuators 10First vehicle 11Further vehicle 12a, 12bPotential trajectories; Candidates RPfeil 21Procedure steps in mobile unit 22Procedure steps in central control unit S1Automatic travel S2Detection of a weakness S3Generation and transmission of a demand message S4Receipt and evaluation of demand messages S5Generation and transmission of a recording order S6Receipt of the recording order and recording of sensor data sets S7Generation and transmission of transmission data S8Receipt of the transmission data S9Generation of an updated control model S10Transfer of the updated control model S11Receipt of an updated control model and automatic travel
Claims
1. Method for updating a control model for automatically controlling at least one mobile unit (1) with an automatic driving function, in which a central control unit (2) generates a capture task and transmits it to the mobile unit (1); wherein the mobile unit (1) comprises sensors (3a; 3b; 3c; 3d) and the capture task comprises conditions for capturing sensor data sets by means of the sensors (3a; 3b; 3c; 3d), wherein these conditions define which types of sensor data are intended to be captured in which driving situations; the mobile unit (1) captures the sensor data sets on the basis of the capture task by means of the sensors (3a; 3b; 3c; 3d), generates transmission data on the basis of the captured sensor data sets, and transmits the transmission data to the central control unit (2); and the central control unit (2) receives the transmission data and, on the basis of the received transmission data, generates an updated control model for automatically controlling the mobile unit (1), wherein, in an exploratory learning phase, a multiplicity of candidates for trajectories (12a, 12b) for automatically controlling the mobile unit (1) are generated, and a trajectory (12a, 12b) that is determined to be safe to use and is not preferred is used, wherein the trajectory (12a, 12b) that is not preferred is reassessed on the basis of the sensor data sets captured during execution, the new assessment results generated in this manner are transmitted to the central control unit (2), and the control model is optimized for automatic journey planning.
2. Method according to Claim 1, wherein a requirement message is received by the central control unit (2) and the capture task is generated on the basis of the requirement message.
3. Method according to Claim 2, wherein the requirement message is generated by the mobile unit (1).
4. Method according to Claim 2 or 3, wherein the requirement message is generated by a traffic monitoring device (8).
5. Method according to one of Claims 2 to 4, wherein the central control unit (2) receives a multiplicity of requirement messages, weights the requirement messages and generates the capture task on the basis of the weighted requirement messages.
6. Method according to one of the preceding claims, wherein the conditions of the capture task for capturing sensor data sets comprise data relating to traffic situations.
7. Method according to one of the preceding claims, wherein the central control unit (2) carries out training for a previous control model in order to generate the updated control model on the basis of the transmission data.
8. Method according to Claim 7, wherein an integrated control model is trained in order to generate the transmission data by means of the mobile unit (1).
9. Method according to one of the preceding claims, wherein the updated control model is transmitted to the mobile unit (1); and the mobile unit (1) is automatically controlled on the basis of the updated control model.
10. System for updating a control model for automatically controlling at least one mobile unit (1) with an automatic driving function, comprising: a central control unit (2) which is designed to generate a capture task and transmit it to the mobile unit (1); wherein the mobile unit (1) comprises sensors (3a; 3b; 3c; 3d) and the capture task comprises conditions for capturing sensor data sets by means of the sensors (3a; 3b; 3c; 3d); wherein the conditions define which types of sensor data are intended to be captured in which driving situations, the sensors (3a; 3b; 3c; 3d) of the mobile unit (1) are designed to capture the sensor data sets on the basis of the capture task, to generate transmission data on the basis of the captured sensor data sets, and to transmit the transmission data to the central control unit (2); wherein the central control unit (2) is designed to receive the transmission data and, on the basis of the received transmission data, to generate an updated control model for automatically controlling the mobile unit (1), wherein the system is also configured such that a multiplicity of candidates for trajectories (12a, 12b) for automatic control are generated, and a trajectory (12a, 12b) that is determined to be safe to use and is not preferred is used, wherein the trajectory (12a, 12b) that is not preferred is reassessed on the basis of the sensor data sets captured during execution, the new assessment results generated in this manner are transmitted to the central control unit (2), and the control model is optimized for automatic journey planning.
