Autonomous interaction model and training method for interacting with off-board personnel

Abstract

The present disclosure provides an autonomous driving model and a training method capable of autonomous interaction with an off-vehicle person. The present disclosure relates to the technical field of autonomous driving. The autonomous driving model comprises a multi-modal encoding layer and a decision control layer for connecting to constitute an end-to-end neural network model, the decision control layer obtains autonomous driving strategy information based on the output of the multi-modal encoding layer, the first input information of the multi-modal encoding layer comprises navigation information of the vehicle and perception information of the environment around the vehicle obtained by using a sensor, the multi-modal encoding layer is used to obtain a hidden representation corresponding to the first input information, and the decision control layer is used to obtain the autonomous driving strategy information based on at least the hidden representation of the interaction information and the perception information. The autonomous driving model can determine the autonomous driving strategy information based on the surrounding environment and the content of the interaction information, so that the driving behavior of the vehicle can adapt to the requirements of the person outside the vehicle.

Description

TECHNICAL FIELD

[0001] The present disclosure relates to the technical field of computer, in particular to the technical field of automatic driving and artificial intelligence, and more particularly to an automatic driving model capable of autonomous interaction with an out-of-vehicle person, an automatic driving method implemented by using the automatic driving model, a training method of the automatic driving model, an automatic driving device based on the automatic driving model, a training device of the automatic driving model, an electronic device, a computer readable storage medium, a computer program product and an automatic driving vehicle. BACKGROUND

[0002] Artificial intelligence is a discipline that studies enabling computers to simulate some thinking processes and intelligent behaviors (such as learning, reasoning, thinking, planning, etc.) of human beings, which has both hardware technology and software technology. Artificial intelligence hardware technology generally includes technologies such as sensors, special artificial intelligence chips, cloud computing, distributed storage, big data processing, etc.; artificial intelligence software technology mainly includes computer vision technology, speech recognition technology, natural language processing technology, and machine learning / deep learning, big data processing technology, knowledge graph technology, etc.

[0003] Automatic driving technology integrates technologies such as recognition, decision-making, positioning, communication safety and human-computer interaction. Through artificial intelligence learning, it can assist in generating automatic driving strategies.

[0004] High-precision map, also known as high-precision map, is a map used by autonomous vehicles. High-precision map has accurate vehicle position information and rich road element data information, which can help vehicles predict complex road information such as slope, curvature, heading, etc., so as to better avoid potential risks. In other words, automatic driving technology strongly depends on high-precision map.

[0005] The methods described in this section are not necessarily the methods that have been previously conceived or employed. Unless otherwise indicated, it should not be assumed that any of the methods described in this section qualify as prior art merely by virtue of their inclusion in this section. Similarly, issues mentioned in this section should not be assumed to have been admitted to be prior art against anyone. SUMMARY

[0006] The present disclosure provides an automatic driving model, an automatic driving method implemented by using the automatic driving model, a training method of the automatic driving model, an automatic driving device based on the automatic driving model, a training device of the automatic driving model, an electronic device, a computer readable storage medium, a computer program product and an automatic driving vehicle.

[0007] According to an aspect of the present disclosure, there is provided an automatic driving model comprising a multi-modal encoding layer and a decision control layer connected to form an end-to-end neural network model, such that the decision control layer obtains automatic driving policy information based on an output of the multi-modal encoding layer, wherein first input information of the multi-modal encoding layer comprises navigation information of a target vehicle and perception information of a surrounding environment of the target vehicle obtained by a sensor, the perception information comprising current perception information and historical perception information of the surrounding environment of the target vehicle in a driving process of the vehicle, the multi-modal encoding layer is configured to obtain a latent representation corresponding to the first input information, second input information of the decision control layer comprises the latent representation, the decision control layer is configured to obtain target automatic driving policy information based on the second input information, and the second input information comprises interaction information of a person outside the target vehicle.

[0008] According to another aspect of the present disclosure, there is provided an automatic driving method implemented by an automatic driving model comprising a multi-modal encoding layer and a decision control layer connected to form an end-to-end neural network model, such that the decision control layer obtains automatic driving policy information based on an output of the multi-modal encoding layer. The method comprises: obtaining first input information of the multi-modal encoding layer, the first input information comprising navigation information of a target vehicle and perception information of a surrounding environment of the target vehicle obtained by a sensor, the perception information comprising current perception information and historical perception information of the surrounding environment of the target vehicle in a driving process of the vehicle; inputting the first input information into the multi-modal encoding layer to obtain a latent representation corresponding to the first input information output by the multi-modal encoding layer; and inputting second input information comprising the latent representation into the decision control layer to obtain target automatic driving policy information output by the decision control layer, wherein the second input information comprises interaction information of a person outside the target vehicle.

[0009] According to another aspect of the present disclosure, a training method of an autonomous driving model is provided, the autonomous driving model comprising a multi-modal encoding layer and a decision control layer, the multi-modal encoding layer and the decision control layer being connected to form an end-to-end neural network base model, such that the decision control layer obtains autonomous driving policy information based on an output of the multi-modal encoding layer, the training method comprising: obtaining sample input information and real autonomous driving policy information corresponding to the sample input information, the sample input information comprising sample navigation information of a sample vehicle and sample perception information for a surrounding environment of the sample vehicle, the sample perception information comprising current sample perception information and historical sample perception information for the surrounding environment of the sample vehicle; inputting the sample input information into the multi-modal encoding layer to obtain a sample implicit representation output by the multi-modal encoding layer; inputting intermediate sample input information comprising the sample implicit representation into the decision control layer to obtain predicted autonomous driving policy information output by the decision control layer, wherein the intermediate sample input information comprises sample interaction information of a person outside the sample vehicle; and adjusting parameters of the multi-modal encoding layer and the decision control layer based at least on the predicted autonomous driving policy information and the real autonomous driving policy information.

[0010] According to another aspect of the present disclosure, an autonomous driving device based on an autonomous driving model is provided, the autonomous driving model comprising a multi-modal encoding layer and a decision control layer, the multi-modal encoding layer and the decision control layer being connected to form an end-to-end neural network model, such that the decision control layer obtains autonomous driving policy information based on an output of the multi-modal encoding layer. The device comprises: an input information obtaining unit configured to obtain first input information of the multi-modal encoding layer, the first input information comprising navigation information of a target vehicle and perception information for a surrounding environment of the target vehicle obtained by a sensor, the perception information comprising current perception information and historical perception information for the surrounding environment of the target vehicle during driving of the vehicle; a multi-modal encoding unit configured to input the first input information into the multi-modal encoding layer to obtain an implicit representation corresponding to the first input information output by the multi-modal encoding layer; and a decision control unit configured to input second input information comprising the implicit representation into the decision control layer to obtain target autonomous driving policy information output by the decision control layer, wherein the second input information comprises interaction information of a person outside the target vehicle.

[0011] According to another aspect of the present disclosure, there is provided a device for training an autonomous driving model, the autonomous driving model comprising a multi-modal encoding layer and a decision control layer connected to form an end-to-end neural network model, such that the decision control layer obtains autonomous driving policy information based on an output of the multi-modal encoding layer, the device being configured to train the multi-modal encoding layer and the decision control layer. The device comprises: a sample information obtaining unit configured to obtain sample input information and real autonomous driving policy information corresponding to the sample input information, the sample input information comprising sample navigation information of a sample vehicle and sample perception information for a surrounding environment of the sample vehicle, the sample perception information comprising current sample perception information and historical sample perception information for the surrounding environment of the sample vehicle; a multi-modal encoding layer training unit configured to input the sample input information into the multi-modal encoding layer to obtain sample implicit representation output by the multi-modal encoding layer; a decision control layer training unit configured to input intermediate sample input information comprising the sample implicit representation into the decision control layer to obtain predicted autonomous driving policy information output by the decision control layer, wherein the intermediate sample input information comprises sample interaction information of a person outside the sample vehicle; and a parameter adjusting unit configured to adjust parameters of the multi-modal encoding layer and the decision control layer based at least on the predicted autonomous driving policy information and the real autonomous driving policy information.

[0012] According to another aspect of the present disclosure, there is provided an electronic device comprising: at least one processor; and a memory communicatively connected to the at least one processor; the memory storing instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the method described above.

[0013] According to another aspect of the present disclosure, there is provided a non-transitory computer-readable storage medium storing computer instructions, wherein the computer instructions are used to cause the computer to perform the method described above.

[0014] According to another aspect of the present disclosure, there is provided a computer program product comprising a computer program, wherein the computer program, when executed by a processor, implements the method described above.

[0015] According to another aspect of the present disclosure, there is provided an autonomous driving vehicle comprising one of the autonomous driving device, the device for training an autonomous driving model, and the electronic device according to embodiments of the present disclosure.

[0016] According to embodiments of the present disclosure, the autonomous driving model can determine autonomous driving policy information based on the content of the surrounding environment and the interaction information, so that the driving behavior of the vehicle can adapt to the requirements of the person outside the vehicle.

[0017] It is to be understood that the details set forth herein do not limit the scope of the embodiments of the present disclosure but merely constitute illustrative examples of how the embodiments can be BRIEF DESCRIPTION OF DRAWINGS

[0018] The accompanying drawings, which are included to provide a further understanding of the application and are incorporated in

[0019] Figure 1 shows a schematic diagram of an exemplary system in which various methods described herein can be implemented, according to an embodiment of the present disclosure;

[0020] Figure 2 shows a schematic diagram of an autonomous driving model, according to an embodiment of the present disclosure;

[0021] Figure 3 shows a flowchart of an autonomous driving method implemented using an autonomous driving model, according to an embodiment of the present disclosure;

[0022] Figure 4 shows a flowchart of a training method of an autonomous driving model, according to an embodiment of the present disclosure;

[0023] Figure 5 shows a structural block diagram of an autonomous driving apparatus based on an autonomous driving model, according to an embodiment of the present disclosure;

[0024] Figure 6 shows a structural block diagram of a training apparatus of an autonomous driving model, according to an embodiment of the present disclosure; and

[0025] Figure 7 shows a structural block diagram of an exemplary electronic device that can be used to implement an embodiment of the present disclosure. DETAILED DESCRIPTION

[0026] Exemplary embodiments of the present disclosure are described herein with reference to the accompanying drawings to assist in a comprehensive understanding of the present disclosure by those of ordinary skill in the art, which include various details of the embodiments of the present disclosure. However, it should be understood that the embodiments of the present disclosure can be practiced with

[0027] In the present disclosure, the terms "first", "second", and the like are used to describe various elements only for the purpose of describing particular examples, and are not intended to limit the positional relationship, the timing relationship, or the importance relationship of the elements. In some examples, the first element and the second element can refer to the same instance of the element, and in some cases, based on the context of the description, they can also refer to different instances.

[0028] The terms used in the description of various described examples in the present disclosure are only for the purpose of describing the particular examples, and are not intended to be limiting. Unless the context clearly indicates otherwise, if the number of elements is not specifically limited, the element can be one or more. In addition, the term "and / or" used in the present disclosure encompasses any one of the listed items and all possible combinations thereof.

[0029] In the technical solutions of the present disclosure, the acquisition, storage and application of user personal information comply with relevant laws and regulations and do not violate public order and good customs.

[0030] In the related art, the algorithm mainly based on optimization and rules in the automatic driving technology usually relies on high-precision maps and algorithm optimization for different scenarios. High-precision map, also known as high-precision map, mainly includes two types of information: one is road information, including the position, category, width, slope and curvature of the lane of the highway and the like; the other is the information of the auxiliary facilities and structures related to the lane, including traffic signs, traffic signal lights, overpass, traffic monitoring points (electronic eyes, speed measurement radars), roadside facilities, obstacles and other road details and infrastructure information, including lane restriction scenarios (such as lane restriction at a certain period) and lane restriction information (such as vehicle type, weather condition, passing time) and the like. Through these data, the navigation system of the automatic driving vehicle can complete accurate positioning, judge which roads can be driven, and provide guidance for the vehicle.

[0031] In the related art, the unmanned technology mainly relies on the cooperation of the perception module and the planning control module. The working process of automatic driving includes two stages: first, the unstructured information obtained by sensors such as cameras or radars is converted into structured information (the structured information includes obstacle information, other vehicle information, pedestrian and non-motor vehicle information, lane line information, traffic light information, other static road surface information, etc.). These information can be matched with high-precision map to accurately obtain the position information on the high-precision map. Second, prediction and decision are made based on the structured information and the related observation history. Among them, the prediction includes predicting the changes of the surrounding structured environment in a period of time in the future; the decision includes generating some structured information that can be used for subsequent trajectory planning (such as lane changing and overtaking, waiting). Third, based on the structured decision information and the changes of the surrounding structured environment, the trajectory of the target vehicle in a period of time in the future is planned, such as planning trajectory or control information (such as planning speed and position).

[0032] It is found through research that the automatic driving technology based on perception-prediction-planning may face some technical problems. First, the problem of error accumulation, because perception cannot be directly responsible for decision-making, which makes perception not necessarily able to capture information that plays a key role in decision-making, in addition, because the errors of perception are difficult to be compensated in the subsequent process (for example, obstacles in the area may not be identified), the subsequent process may be difficult to make correct decisions in the case of missing key obstacles. Second, the coupling problem between prediction and planning cannot be solved, the behavior of surrounding obstacles, especially the key obstacles interacting with the target vehicle, may be affected by the target vehicle. In other words, during the running process of the automatic driving model, there is a coupling between the prediction and planning modules, which makes the streaming decision affect the final automatic driving effect. In addition, there is a problem of representation defect of structured information, structured information is completely limited to artificially predefined standards, once encountering new paradigms that are not clearly defined (such as unknown obstacles, unknown states of vehicles and pedestrians, etc.), the algorithm is easy to fail. Finally, there is a problem of dependence on high-cost maps (such as high-precision maps), the related art mainly relies on high-precision map point cloud and other information to perform vehicle positioning, however, in practice, high-precision maps are only available in limited areas, which limits the actual application area of automatic driving; in addition, the updating cost of high-precision maps is huge, once the map and the actual road do not match, it is easy to cause decision failure.

[0033] Based on this, the present disclosure provides an automatic driving model, an automatic driving method implemented by using the automatic driving model, a training method of the automatic driving model, an automatic driving device based on the automatic driving model, a training device of the automatic driving model, an electronic device, a computer readable storage medium, a computer program product and an automatic driving vehicle, an automatic driving technology of perception-decision integration is adopted, so that the perception is directly responsible for the decision, which is beneficial to the perception to capture information that plays a key role in the decision, reduces error accumulation, and solves the coupling problem between prediction and decision in the related art. In addition, the perception directly responsible for the decision can also overcome the problem that the structured information is limited by the artificially predefined standard, so that the algorithm is prone to failure, and realize the automatic driving technology of heavy perception and light map, thereby being able to overcome the problem that the high-precision map update is not timely and the area is limited, resulting in the failure of the decision. Since the dependence on the high-precision map is eliminated, the update cost of the high-precision map can be saved.

[0034] Embodiments of the present disclosure will be described in detail below with reference to the accompanying drawings.

[0035] Figure 1 A schematic diagram of an example system 100 in which various methods and apparatus described herein can be implemented according to embodiments of the present disclosure is shown. Referring to Figure 1 The system 100 includes a motor vehicle 110, a server 120, and one or more communication networks 130 coupling the motor vehicle 110 to the server 120.

[0036] In embodiments of the present disclosure, the motor vehicle 110 can include a computing device according to embodiments of the present disclosure and / or be configured to perform methods according to embodiments of the present disclosure.

[0037] The server 120 can run one or more services or software applications that enable autonomous driving. In certain embodiments, the server 120 can also provide other services or software applications, which can include non-virtual environments and virtual environments. In Figure 1 In the illustrated configuration, the server 120 can include one or more components implementing the functionality performed by the server 120. These components can include software components, hardware components, or a combination thereof, executable by one or more processors. Users of the motor vehicle 110 can in turn utilize one or more client applications to interact with the server 120 to utilize the services provided by these components. It should be understood that various different system configurations are possible, which can differ from the system 100. Thus, Figure 1 The system 100 is one example of a system for implementing the various methods described herein and is not intended to be limiting.

[0038] The server 120 can include one or more general purpose computers, special purpose server computers (e.g., PC (personal computer) servers, UNIX servers, mid-range servers), blade servers, mainframe computers, server clusters, or any other appropriate arrangement and / or combination. The server 120 can include one or more virtual machines running virtual operating systems, or other computing architectures involving virtualization (e.g., one or more flexible pools of logical storage devices that can be virtualized to maintain virtual storage devices for the server). In various embodiments, the server 120 can run one or more services or software applications that provide the functionality described below.

[0039] The computing units in the server 120 can run one or more operating systems including any of the operating systems described above, as well as any commercially available server operating systems. The server 120 can also run any of a variety of additional server applications and / or mid-tier applications, including HTTP servers, FTP servers, CGI servers, JAVA servers, database servers, etc.

[0040] In some implementations, the server 120 can include one or more applications to analyze and consolidate data feeds and / or event updates received from the motor vehicles 110. The server 120 can also include one or more applications to display the data feeds and / or real-time events via one or more display devices of the motor vehicles 110.

[0041] The network 130 can be any type of network familiar to those skilled in the art that can support data communications using any of a variety of available protocols, including without limitation TCP / IP, SNA, IPX, etc. As an example, one or more of the networks 130 can be a satellite communications network, a local area network (LAN), an Ethernet-based network, a token ring, a wide area network (WAN), the Internet, a virtual network, a virtual private network (VPN), an intranet, an extranet, a blockchain network, a public switched telephone network (PSTN), an infrared network, a wireless network (including, for example, a Bluetooth, 802.11, etc.), and / or any combination of these and other networks.

[0042] The system 100 can also include one or more databases 150. In certain embodiments, these databases can be used to store data and other information. For example, one or more of the databases 150 can be used to store information such as audio files and video files. The data stores 150 can reside in various locations. For example, a data store used by the server 120 can be local to the server 120, or can be remote from the server 120 and can communicate with the server 120 via a network- or dedicated-based connection. The data stores 150 can be of different types. In certain embodiments, a data store used by the server 120 can be a database, such as a relational database. One or more of these databases can store, update, and retrieve data to and from the database in response to commands.

[0043] In certain embodiments, one or more of the databases 150 can also be used by applications to store application data. Databases used by applications can be databases of different types, such as key-value stores, object stores, or regular stores backed by file systems.

[0044] The motor vehicle 110 can include sensors 111 for perceiving the surrounding environment. The sensors 111 can include one or more of the following sensors: visual camera, infrared camera, ultrasonic sensor, millimeter wave radar, and laser radar (LiDAR). Different sensors can provide different detection accuracy and range. The camera can be installed at the front, rear, or other positions of the vehicle. The visual camera can capture the situation inside and outside the vehicle in real time and present it to the driver and / or passenger. In addition, through analysis of the pictures captured by the visual camera, information such as traffic signal indication, intersection situation, running state of other vehicles, etc. can be obtained. The infrared camera can capture objects in night vision conditions. The ultrasonic sensor can be installed around the vehicle to measure the distance of objects outside the vehicle from the vehicle by taking advantage of the strong directivity of ultrasonic waves. The millimeter wave radar can be installed at the front, rear, or other positions of the vehicle to measure the distance of objects outside the vehicle from the vehicle by taking advantage of the characteristics of electromagnetic waves. The laser radar can be installed at the front, rear, or other positions of the vehicle to detect object edges, shape information, and thus perform object recognition and tracking. Due to the Doppler effect, the radar device can also measure the speed change of the vehicle and moving objects.

[0045] The motor vehicle 110 can further comprise a communication device 112. The communication device 112 can comprise a satellite positioning module capable of receiving satellite positioning signals (e.g. Beidou, GPS, GLONASS and GALILEO) from satellites 141 and generating coordinates based on these signals. The communication device 112 can further comprise a module for communicating with mobile communication base stations 142, which can implement any suitable communication technology, such as GSM / GPRS, CDMA, LTE, etc. current or evolving wireless communication technologies (e.g. 5G technology). The communication device 112 can further have a vehicle-to-everything (V2X) module configured for enabling vehicle-to-vehicle (V2V) communication with other vehicles 143 and vehicle-to-infrastructure (V2I) communication with infrastructure 144, for example. In addition, the communication device 112 can further have a module configured to communicate with user terminals 145 (including but not limited to smartphones, tablets or wearable devices such as watches) for example through a wireless local area network using IEEE 802.11 standards or Bluetooth. With the communication device 112, the motor vehicle 110 can further access the server 120 via the network 130.

[0046] The motor vehicle 110 can further comprise a control device 113. The control device 113 can comprise a processor, such as a central processing unit (CPU) or a graphics processing unit (GPU), or other special purpose processor, in communication with various types of computer readable storage devices or media. The control device 113 can comprise an autonomous driving system for automatically controlling various actuators in the vehicle. The autonomous driving system is configured to control the powertrain, steering system and braking system of the motor vehicle 110 (not shown) via a plurality of actuators to control acceleration, steering and braking, respectively, in response to inputs from a plurality of sensors 111 or other input devices, without or with limited human intervention. Part of the processing functions of the control device 113 can be implemented through cloud computing. For example, some processing can be performed using an on-board processor, while other processing can be performed using computing resources in the cloud. The control device 113 can be configured to perform methods according to the present disclosure. In addition, the control device 113 can be implemented as an example of a motor vehicle-side (client-side) computing device according to the present disclosure.

[0047] Figure 1 The system 100 can be configured and operated in various ways to enable the application of various methods and devices described according to the present disclosure.

[0048] According to an aspect of the present disclosure, an autonomous driving model is provided.Figure 2 A schematic diagram of an autonomous driving model 200 is shown according to an embodiment of the present disclosure.

[0049] As shown, the autonomous driving model 200 includes a multi-modal encoding layer 210 and a decision control layer 220, which are connected to form an end-to-end neural network model, so that the decision control layer 220 obtains autonomous driving policy information based on the output of the multi-modal encoding layer 210. The first input information of the multi-modal encoding layer 210 includes navigation information In1 of a target vehicle and perception information of the surrounding environment of the target vehicle obtained by using sensors (for example, but not limited to, In2, In3 and In4, which are taken as examples in the following description of the perception information including In2, In3 and In4), and the perception information includes current perception information and historical perception information of the surrounding environment of the target vehicle during the driving process of the target vehicle. The multi-modal encoding layer 210 is configured to obtain an implicit representation e t based on the first input information In1-In4. t The second input information of the decision control layer 220 includes the implicit representation e t , and the decision control layer 220 is configured to obtain target autonomous driving policy information based on the second input information.

[0050] As described above, in the related art, prediction can be performed based on the perception information to obtain future prediction information, and the decision control layer performs planning based on the future prediction information, that is, the decision control layer 220 does not directly perform planning based on the perception information, but directly performs planning based on the future prediction information. In the embodiment of the present application, the decision control layer 220 can directly obtain autonomous driving policy information based on the output of the multi-modal encoding layer 210, and the multi-modal encoding layer 210 is used to encode and calculate the perception information, which is equivalent to that the decision control layer 220 can directly perform planning based on the perception information to obtain autonomous driving policy information. In other words, in the embodiment of the present application, the perception is directly responsible for decision-making.

[0051] In an example, the autonomous driving model 200 can adopt a Transformer network structure with an encoder (Encoder) and a decoder (Decoder). It can be understood that the autonomous driving model 200 can also be other neural network models based on the Transformer network structure, which are not limited herein. The Transformer architecture can calculate the implicit representation of the input and output of the model through a self-attention mechanism. In other words, the Transformer architecture can be an Encoder-Decoder model constructed based on such a self-attention mechanism.

[0052] In an example, the navigation information In1 of the target vehicle in the first input information can include vectorized navigation information and vectorized map information, which can be obtained by vectorizing one or more of lane-level or road-level navigation information and coarse positioning information.

[0053] According to some embodiments of the present application, the perception information In2, In3 and In4 of the surrounding environment of the target vehicle can include perception information In2 of one or more cameras, perception information In3 of one or more lidars, and perception information In4 of one or more millimeter wave radars. It can be understood that the perception information of the surrounding environment of the target vehicle is not limited to the above-mentioned form, for example, it can only include perception information In2 of multiple cameras, and does not include perception information In3 of one or more lidars and perception information In4 of one or more millimeter wave radars. The perception information In2 obtained by the camera can be picture or video form perception information, and the perception information In3 obtained by the lidar can be radar point cloud (such as three-dimensional point cloud) form perception information. In an example, the above-mentioned different forms of information (picture, video, point cloud) and the like can be directly input to the multi-modal encoding layer 210 without pre-processing. In addition, the perception information includes current perception information x t and historical perception information x t-Δt corresponding to multiple historical time points, where t and Δt can have a time span of a preset time length.

[0054] In an example, the multi-modal encoding layer 210 can perform encoding calculation on the first input information to generate corresponding implicit representation e t . The implicit representation e t may be, for example, an implicit representation in the bird's eye view (BEV) space. For example, the perception information In2 of the camera can be first input to a shared backbone network to extract data features of each camera. Then, the perception information In2 of multiple cameras is fused and converted to the BEV space. Then, cross-modal fusion can be performed in the BEV space to fuse pixel-level visual data and lidar point cloud. Finally, temporal fusion is performed to form the implicit representation e t in the BEV space.

[0055] In an example, a Transformer Encoder structure that fuses spatio-temporal information can be used to realize the implicit representation e tthe projection. For example, the spatio-temporal information can be utilized by a BEV query mechanism with grid partition of pre-set parameters. By utilizing a spatial cross-attention mechanism (i.e., the BEV query mechanism extracts the required spatial features from the multi-camera features by the attention mechanism), the BEV query mechanism is enabled to extract features from the multi-camera perspectives of interest thereof, thereby aggregating the spatial information; in addition, by utilizing a temporal self-attention mechanism (i.e., the BEV features generated at each time instant obtain the required temporal information from the BEV features of the previous time instant), the historical information is fused, thereby aggregating the temporal information.

[0056] Correspondingly, the decision control layer 220 obtains target autonomous driving policy information based on the input implicit representation e t For example, the target autonomous driving policy information can include a planned trajectory Out1 or a control signal Out2 for the vehicle (e.g., a signal for controlling the throttle, brake, steering amplitude, etc.). In an example, the trajectory planning Out1 can be interpreted by a control policy module in the autonomous vehicle to obtain the control signal Out2 for the vehicle; or a neural network can be utilized to directly output the control signal Out2 for the vehicle based on the implicit representation e t .

[0057] In an example, the decision control layer 220 can include a decoder in the Transformer.

[0058] In Figure 2 , the solid arrows between the multi-modal encoding layer 210 to the decision control layer 220 and the decision control layer 220 to the trajectory planning Out1 represent differentiable operations, in other words, when the model is trained, the gradient can be back-propagated through the above-mentioned solid arrows.

[0059] It can be seen that in the autonomous driving model 200 according to the embodiments of the present disclosure, the multi-modal encoding layer 210 and the decision control layer 220 are connected to constitute an end-to-end neural network model, so that the perception information can be directly responsible for the decision, and the coupling problem between prediction and planning can be solved. In addition, the introduction of the implicit representation can overcome the problem that the structured information is prone to failure. In addition, since the perception is directly responsible for the decision, the perception can capture information that is more critical to the decision, reducing the accumulation of errors caused by perception errors. Furthermore, since the perception is directly responsible for the decision, the autonomous driving technology of heavy perception and light map is realized, thereby overcoming the problem of decision failure caused by the update of high-precision maps not being timely and the area being limited. Since the dependence on high-precision maps is eliminated, the update cost of high-precision maps can be saved.

[0060] According to some embodiments, continuing to refer to Figure 2, the autonomous driving model 200 can further comprise a future prediction layer 230 configured for predicting future prediction information Out3 for the target vehicle's surrounding environment based on the input implicit representation e t The future prediction information Out3 can be predicted by the future prediction layer 230 based on the input implicit representation e t The predicted obstacle position at a future time or the sensor input information at a future time. At least part of the future prediction information Out3 can be input into the decision control layer 220 as auxiliary information A, and the decision control layer 220 can predict the target autonomous driving strategy information based on the implicit representation e t and the auxiliary information A.

[0061] In an example, the future prediction layer 230 can be a decoder in a Transformer.

[0062] In an example, the future prediction information Out3 can output structured prediction information, and accordingly, the dashed arrows between the future prediction information Out3 and the auxiliary information A, and the auxiliary information A and the decision control layer 220 represent non-differentiable operations, in other words, the gradient cannot be back-propagated through the above dashed arrows when training the model. However, since the operations between the multi-modal encoding layer 210 and the future prediction layer 230, and the future prediction layer 230 and the future prediction information Out3 are differentiable, the gradient can still be back-propagated in the direction indicated by the solid arrows, in other words, the future prediction layer 230 can also be trained separately.

[0063] Therefore, by introducing the future prediction layer 230 in the autonomous driving model 200, at least part of the information predicted by the future prediction layer 230 is input into the decision control layer 220 as auxiliary information to assist decision-making, which can improve the accuracy and safety of decision-making. In addition, when training the model, the multi-modal encoding layer 210 can be further trained by the future prediction layer 230 based on the decision control layer 220, so that the encoding of the multi-modal encoding layer 210 is more accurate, and thus the decision control layer 220 can predict more optimized target autonomous driving strategy information.

[0064] According to some embodiments, the future prediction information Out3 can comprise at least one of: future prediction perception information (e.g., sensor information at a future time for the target vehicle's surrounding environment, future prediction implicit representation corresponding to the future prediction perception information (for example, an implicit representation in the BEV space corresponding to sensor information at a future time), and future prediction detection information (for example, obstacle positions at a future time ) for the target vehicle’s surrounding environment. Moreover, the future prediction detection information can include types of multiple obstacles in the target vehicle’s surrounding environment and their future prediction state information (including sizes of the obstacles and various long-tail information).

[0065] According to some embodiments, continuing to refer to Figure 2 , the autonomous driving model 200 can further include a perception detection layer 240, which can be configured to obtain target detection information Out4 for the target vehicle’s surrounding environment based on the input implicit representation e t The target detection information Out4 includes current detection information and historical detection information, the current detection information including types of multiple road elements and obstacles in the target vehicle’s surrounding environment and their current state information, and the historical detection information including types of multiple obstacles in the target vehicle’s surrounding environment and their historical state information. And the second input information of the decision control layer 220 can further include at least a portion of the target detection information Out4.

[0066] The road elements can be static objects, while the obstacles can be moving objects, so the historical state information of the road elements can not be detected.

[0067] In an example, the target detection information Out4 can be a bounding box in a three-dimensional space for an obstacle, and can indicate a classification, a state, etc. of the corresponding obstacle in the bounding box. For example, the size, the position, and the vehicle type, the current state (for example, whether the turn signal, the high beam, etc. long-tail information is on) of the obstacle in the bounding box, the position and the length of the lane line, etc. can be indicated. It will be understood that the classification for the corresponding obstacle in the bounding box can be one or more of a plurality of predefined categories.

[0068] In addition, the target detection information Out4 (the current detection information and the historical detection information) can be structured information. Accordingly, the dashed arrows between the target detection information Out4 and the auxiliary information A, the auxiliary information A and the decision control layer 220 represent non-differentiable operations, in other words, the gradients can not be back-propagated through the above dashed arrows when training the model. However, since the operations between the multi-modal encoding layer 210 and the perception detection layer 240, and the perception detection layer 240 and the target detection information Out4 are differentiable, the gradients can still be back-propagated in the direction indicated by the solid arrows, in other words, the perception detection layer 240 can also be trained separately.

[0069] In an example, the perception detection layer 240 can include a decoder in a Transformer.

[0070] Thus, by introducing the perception detection layer 240 in the autonomous driving model 200, at least part of the information predicted by the perception detection layer 240 is input into the decision control layer 220 as auxiliary information to assist decision making, which can make the detection information of the current and historical period of time of the surrounding environment of the vehicle available for assisting decision making, thereby improving the accuracy and safety of decision making. In addition, when training the model, the multi-modal encoding layer 210 can be further trained by the perception detection layer 240 on the basis of the decision control layer 220, so that the encoding of the multi-modal encoding layer 210 is more accurate, so that the decision control layer 220 can predict more optimized target autonomous driving strategy information.

[0071] According to some embodiments, with continued reference to Figure 2 , the autonomous driving model 200 can further include an evaluation feedback layer 250, which can be configured to input the implicit representation e t obtain evaluation feedback information Out5 for the target autonomous driving strategy information.

[0072] In an example, the evaluation feedback layer 250 can be a decoder in the Transformer.

[0073] Thus, by introducing the evaluation feedback layer 250 in the autonomous driving model 200, it can be indicated whether the current driving behavior is derived from a human driver or a model, whether the current driving is comfortable, whether the current driving violates traffic rules, and whether the current driving is dangerous driving, etc., thereby improving user experience.

[0074] It will be understood that the solid arrows between the multi-modal encoding layer 210 to the evaluation feedback layer 250, the evaluation feedback layer 250 to the evaluation feedback information Out5 represent differentiable operations, in other words, when training the model, the gradient can be backpropagated through the above solid arrows. Thus, when training the model, the multi-modal encoding layer 210 can be further trained by the evaluation feedback layer 250 on the basis of the decision control layer 220, so that the encoding of the multi-modal encoding layer 210 is more accurate, so that the decision control layer 220 can predict more optimized target autonomous driving strategy information.

[0075] According to some embodiments, as Figure 2 the auxiliary information A including the future prediction information Out3 and the target detection information Out4 is input into the evaluation feedback layer 250 as shown by the dashed arrow, when the autonomous driving model 200 includes the future prediction layer 230 and the perception detection layer 240, the evaluation feedback layer 250 can be configured to input at least part of one or both of the future prediction information Out3 and the target detection information Out4, and the implicit representation et The evaluation feedback information Out5 for the target autonomous driving policy information is obtained. In this way, the current and historical detection information and the future prediction information of the vehicle surrounding environment can be used to assist the evaluation, improving the accuracy of the evaluation.

[0076] According to some embodiments, the evaluation feedback layer 250 can be configured to input the implicit representation e t The evaluation feedback information for the target autonomous driving policy information is obtained based on the target autonomous driving policy information (e.g., the planning trajectory Out1). In this way, the evaluation feedback is assisted based on the autonomous driving policy information, which can further improve the accuracy of the evaluation.

[0077] According to some embodiments, the evaluation feedback layer 250 can be configured to input the implicit representation e t The evaluation feedback information Out5 for the target autonomous driving policy information is obtained, which can further improve the accuracy of the evaluation.

[0078] According to some embodiments, further referring to Figure 2 The autonomous driving model 200 can further include an explanation layer 260, which can be configured to input the implicit representation e t The explanation information Out6 for the target autonomous driving policy information is obtained, which can represent the decision classification of the target autonomous driving policy information. In this way, in the autonomous driving process, the passengers can be provided with the explanation information related to the target autonomous driving policy information, improving the explainability of the autonomous driving policy, thereby improving the user experience.

[0079] In an example, the explanation layer 260 can classify the target autonomous driving policy information, and each classification can be mapped to a preset natural language sentence. For example, the explanation information Out6 can include natural language sentences such as: current need to change lanes, there is a traffic light in front so need to slow down, surrounding vehicles may need to cut in, etc. In addition, the explanation layer 260 can include a decoder in the Transformer to decode the natural language for the explanation of the driving policy.

[0080] According to some embodiments, when the autonomous driving model 200 includes the future prediction layer 230 and the perception detection layer 240, the explanation layer 260 can be configured to input the implicit representation e tThe explanation information Out6 for the target autonomous driving strategy information is obtained. In this way, the target detection information and the future prediction information for the current and historical period of time of the vehicle surrounding environment can be used to assist the explanation, so as to further improve the accuracy and rationality of the explanation.

[0081] According to some embodiments, continuing to refer to Figure 2 , the explanation layer 260 can be configured to obtain the explanation information for the target autonomous driving strategy information based on the implicit representation e t and the target autonomous driving strategy information (e.g., the planning trajectory Out1). In this way, the autonomous driving strategy information is used to assist the explanation, which can further improve the accuracy of the explanation.

[0082] According to some other embodiments of the present application, the explanation layer 260 can be configured to obtain at least part of one or both of the future prediction information Out3 and the target detection information Out4 of the input, the target autonomous driving strategy information, and the implicit representation e t The explanation information Out6 for the target autonomous driving strategy information is obtained, so as to further improve the accuracy of the explanation.

[0083] According to some embodiments, the sensor can include a camera, and the perception information can include a two-dimensional image captured by the camera. Moreover, the multi-modal encoding layer 210 can be further configured to obtain the implicit representation e t based on the first input information including the two-dimensional image and the intrinsic and extrinsic parameters of the camera.

[0084] In an example, the intrinsic parameters (i.e., the parameters related to the characteristics of the camera itself, such as the focal length, pixel size, etc. of the camera) and the extrinsic parameters (i.e., the parameters in the world coordinate system, such as the position and rotation direction of the camera) of the camera can be input into the modal encoding layer 210 as the hyperparameters of the autonomous driving model 200. The intrinsic and extrinsic parameters of the camera can be used to convert the input two-dimensional image to, for example, the BEV space.

[0085] In addition, the perception information can be a sequence of two-dimensional images respectively captured by multiple cameras.

[0086] According to some embodiments, the first input information can further include a lane-level map, and the navigation information can include road-level navigation information and / or lane-level navigation information. Unlike the high-definition map, the lane-level map has better availability and smaller space occupation. In this way, by using the lane-level map and the lane-level navigation information, the dependence on the high-definition map can be overcome.

[0087] The navigation map can include a road-level map (SD Map), a lane-level map (LD Map), and a high-definition map (HD Map). The road-level map is mainly composed of road topology information with a certain granularity, and has low navigation positioning accuracy (e.g., the accuracy is about 15 meters or so), and is mainly used to help drivers navigate, and cannot meet the needs of autonomous driving. The lane-level map and the high-definition map can be used for autonomous driving. The lane-level map adds lane-level topology information, has higher accuracy than the road-level map, and generally has sub-meter accuracy, and can include road information (such as lane lines) and auxiliary facilities information related to the lane (such as traffic lights, road signs, parking spaces, etc.), and can be used to assist autonomous driving. Compared with the lane-level map, the high-definition map has higher map data accuracy (the accuracy reaches centimeter level), richer map data types, and higher map update frequency, and can be used for autonomous driving. Among the three kinds of navigation maps, the high-definition map has the richest information and the highest accuracy, and also has higher use and update costs. Since the scheme of the embodiments of the present application directly makes the perception responsible for decision-making, the autonomous driving technology of heavy perception and light map can be realized, so that the dependence on the high-definition map can be eliminated, and efficient decision-making can be ensured. Further, using the lane-level map as auxiliary information for decision-making can improve the decision-making effect.

[0088] According to some embodiments, the perception information can include at least one of: images captured by a camera, information captured by a laser radar, and information captured by a millimeter wave radar. It will be understood that the images obtained by the camera can be in the form of pictures or videos, and the information obtained by the laser radar can be a radar point cloud (e.g., a three-dimensional point cloud).

[0089] According to some embodiments, the multi-modal encoding layer 210 is configured to map the first input information to a preset space to obtain an intermediate representation, and process the intermediate representation using a temporal attention mechanism and / or a spatial attention mechanism to obtain an implicit representation e t .

[0090] In an example, the preset space can be a BEV space. Because the perception, prediction, decision-making, and planning processes are all carried out in a three-dimensional space, the image information captured by the camera is only a projection of the real physical world in a perspective view, and the information obtained from the image needs to be processed in a complex manner before it can be used, so there is a certain loss of information, and mapping the visual information to the BEV space can more conveniently connect the perception and the planning control.

[0091] In an example, the first input information (e.g., image information in the first input information) can be input to a backbone network (e.g., a ResNet, an EfficientNet, or the like) first, and multi-layer image features can be extracted as intermediate representations. In addition, the data of the lidar and the millimeter wave radar can be directly converted to the BEV space. Subsequently, the spatial features required can be extracted from the image features by using a spatial self-attention mechanism, so as to aggregate the spatial information; in addition, the historical information can be fused by using a temporal self-attention mechanism, so as to aggregate the temporal information.

[0092] In this way, the implicit representation e t The rich temporal and spatial information can be represented, so as to further improve the accuracy and safety of the decision.

[0093] According to some embodiments, the target autonomous driving strategy information can include a target planning trajectory Out1.

[0094] In the driving process of the autonomous vehicle, in addition to perceiving the state or movement trajectory of the obstacles and road elements on the road, the autonomous vehicle can also interact with the environment outside the vehicle, and determine the autonomous driving strategy information based on the interaction information from the personnel outside the vehicle. In some embodiments, the second input information of the decision control layer 220 includes the interaction information with the target personnel outside the vehicle. The target personnel outside the vehicle mentioned herein can be personnel who have an impact on the driving behavior of the target vehicle, such as traffic control personnel (e.g., traffic police), pedestrians around the target vehicle, drivers of other motor vehicles or non-motor vehicles around the target vehicle, and the like. The above-mentioned personnel can interact with the target vehicle through language and / or gestures, and the target vehicle can make a decision on the driving strategy according to the interaction information with the above-mentioned personnel.

[0095] In some examples, the interaction information with the personnel outside the target vehicle can include at least one of action interaction information and language interaction information. The action interaction information can include the command gestures of the traffic control personnel, such as stop, continue driving, go straight, turn left, turn right, and the like. The action interaction information can also include the gesture of the pedestrian, such as raising the hand to indicate that the vehicle needs to yield, shaking the hand to indicate that the vehicle does not need to yield, and the like. The language interaction information can include the language of the personnel outside the vehicle, for example, the traffic control personnel, the pedestrian, or the other driver can intervene in the driving behavior of the vehicle through shouting. When the personnel outside the vehicle issues such interaction information, the autonomous driving model can determine the autonomous driving strategy information based on the surrounding environment and the content of the interaction information, so that the driving behavior of the vehicle can adapt to the requirements of the personnel outside the vehicle. The action interaction information can be collected by a camera, and the language interaction information can be collected by a microphone.

[0096] In some implementations, at least part of the interaction information of the person outside the vehicle can be included in the perception information input to the multi-modal encoding layer 210. For example, the action interaction information can be captured by the camera and included in the perception information of the surrounding environment captured by the camera. With the automatic driving model provided by the present disclosure, it is not necessary to separately identify the action interaction information from the obtained image and respond to the identified action interaction information, but the perception information including the interaction information is processed by the multi-modal encoding layer 210 to obtain the implicit representation of the perception information. Among them, the content of the interaction information in the second input information can be included in the implicit representation.

[0097] In other implementations, the first input information can include both the perception information of the surrounding environment and the interaction information of the person outside the target vehicle. The interaction information of the person outside the target vehicle can be input to the multi-modal encoding layer together with the perception information of the surrounding environment of the target vehicle captured by the sensor.

[0098] In some examples, the interaction information can be included in the implicit representation output by the multi-modal encoding layer. For example, in the case where the interaction information is action interaction information captured by the camera, the multi-modal encoding layer can map the image of the action interaction information to an implicit representation, for example, in the BEV space. In other examples, the interaction information can be encoded in other forms different from the implicit representation. In this case, the second input information can include both the implicit representation and the interaction information (or the encoded interaction information). For example, in the case where the interaction information is language interaction information, the multi-modal encoding layer can map the language interaction information to a corresponding token string. The vector form of the token {c1……cn} can be input to the decision control layer together with the implicit representation. In this case, the decision control layer can be implemented as a transformer network structure with coding and decoding integrated, where both coding and decoding use a unidirectional self-attention mechanism. With the above method, the decision control layer can simultaneously consider both the input information related to the perception information (such as the implicit representation corresponding to the perception information) and the language interaction information from the person outside the vehicle to obtain the target automatic driving strategy information. N} with the implicit representation. In this case, the decision control layer can be implemented as a transformer network structure with coding and decoding integrated, where both coding and decoding use a unidirectional self-attention mechanism. With the above method, the decision control layer can simultaneously consider both the input information related to the perception information (such as the implicit representation corresponding to the perception information) and the language interaction information from the person outside the vehicle to obtain the target automatic driving strategy information.

[0099] In some implementations, the interaction information of the person outside the vehicle can also not be included in the first input information. The interaction information can be processed by any other model capable of processing the interaction information to obtain encoded interaction information capable of being processed by the decision control layer. The encoded interaction information, the implicit representation output by the multi-modal encoding layer, and any other auxiliary information are input into the decision control layer as the second input information. In some other implementations, when the decision control layer is configured to be capable of directly processing the interaction information, the interaction information can also be input into the decision control layer as the second input information together with the implicit representation.

[0100] When the second input information includes both the implicit representation and the interaction information, the second input information can further include a time identifier of the interaction information, a type identifier of the interaction information, a time identifier of the implicit representation, and a type identifier of the implicit representation. By assigning different values to the type identifier of the interaction information and the type identifier of the implicit representation, the model can distinguish which part of the input information belongs to the interaction information not included in the perception information and which part belongs to the perception information.

[0101] In some embodiments, the decision control layer 220 can also be configured to generate target interaction information based on the second input information. When the decision control layer 220 is used to output, the last generated vector can be used as input to generate the next vector. For example, after the decision control layer 220 outputs the target automatic driving strategy information, the target automatic driving strategy information can be used as input to generate subsequent output. In this way, interaction content can be generated based on the generated automatic driving strategy information, so that the interaction between the automatic driving model and the user is relevant to the current driving behavior.

[0102] In some implementations, the target interaction information can include an interaction feedback identifier, and when the interaction feedback identifier is true, the target interaction information further includes interaction content. In some examples, the interaction content can include natural language information. Using the above method, the automatic driving model provided by the embodiments of the present disclosure can identify the timing of outputting interaction content and adaptively output target interaction information when needed. Thus, a single model can be used to simultaneously implement the decision of the automatic driving strategy and the interaction with the user, and the historical interaction content of the automatic driving model and the user will help the automatic driving model to implement the decision of the automatic driving strategy.

[0103] According to another aspect of the present disclosure, an automatic driving method implemented by an automatic driving model is provided. Figure 3A flowchart of an autonomous driving method 300 implemented by an autonomous driving model is shown. The autonomous driving model comprises a multi-modal encoding layer and a decision control layer, which are connected to form an end-to-end neural network model, so that the decision control layer obtains autonomous driving strategy information based on the output of the multi-modal encoding layer. For example, the method 300 can be implemented by the autonomous driving model 200 as described above.

[0104] As shown in Figure 3 The autonomous driving method 300 comprises:

[0105] In step S310, first input information of the multi-modal encoding layer is obtained, the first input information comprising navigation information of the target vehicle and perception information of the surrounding environment of the target vehicle obtained by a sensor, the perception information comprising current perception information and historical perception information of the surrounding environment of the target vehicle during driving of the vehicle.

[0106] In step S320, the first input information is input into the multi-modal encoding layer to obtain an implicit representation output by the multi-modal encoding layer corresponding to the first input information.

[0107] In step S330, second input information comprising the implicit representation is input into the decision control layer to obtain target autonomous driving strategy information output by the decision control layer, wherein the second input information comprises interaction information of a person outside the target vehicle.

[0108] By the method provided by the present disclosure, the autonomous driving model can determine autonomous driving strategy information based on the surrounding environment and the content of the interaction information, so that the driving behavior of the vehicle can adapt to the requirements of the person outside the vehicle.

[0109] In some embodiments, the interaction information from outside the target vehicle can comprise at least one of action interaction information and language interaction information.

[0110] In an example, at least part of the interaction information of the person outside the vehicle can be included in the perception information input into the multi-modal encoding layer. Accordingly, the content of the interaction information in the second input information can be included in the implicit representation. In an example, at least part of the interaction information of the person outside the vehicle can be included in the second input information for determining the target autonomous driving strategy information together with the implicit representation.

[0111] In the case where the second input information comprises both the implicit representation and the interaction information, the second input information can further comprise a time identifier of the interaction information, a type identifier of the interaction information, a time identifier of the implicit representation, and a type identifier of the implicit representation.

[0112] In some embodiments, the decision control layer can also be configured to generate target interaction information based on the second input information. In outputting with the decision control layer, the generated last vector can be taken as input to produce the next vector. For example, when the decision control layer outputs target automatic driving strategy information, the target automatic driving strategy information can be taken as input to produce subsequent output.

[0113] In some implementations, the target interaction information can include an interaction feedback identifier, and when the interaction feedback identifier is true, the target interaction information further includes interaction content. In some examples, the interaction content can include natural language information.

[0114] According to another aspect of the present disclosure, a training method of an automatic driving model is provided. Figure 4 A flowchart of the training method of the automatic driving model according to an embodiment of the present disclosure is shown. The automatic driving model includes a multi-modal encoding layer and a decision control layer, which are connected to constitute an end-to-end neural network model, so that the decision control layer obtains automatic driving strategy information based on the output of the multi-modal encoding layer. In examples, the automatic driving model to be trained can adopt a Transformer network structure with an encoder and a decoder. It can be understood that the automatic driving model to be trained can also be other neural network models based on the Transformer network structure, which are not limited herein. For example, the automatic driving model can be the automatic driving model 200 described above.

[0115] As shown in Figure 4 The training method 400 includes:

[0116] Step S410, obtaining sample input information and real automatic driving strategy information corresponding to the sample input information, the sample input information including sample navigation information of a sample vehicle and sample perception information for a surrounding environment of the sample vehicle, the sample perception information including current sample perception information and historical sample perception information for the surrounding environment of the sample vehicle;

[0117] Step S420, inputting the sample input information into the multi-modal encoding layer to obtain sample implicit representation output by the multi-modal encoding layer;

[0118] Step S430, inputting intermediate sample input information including the sample implicit representation into the decision control layer to obtain predicted automatic driving strategy information output by the decision control layer, wherein the intermediate sample input information includes sample interaction information of a person outside the sample vehicle; and

[0119] Step S440, adjusting parameters of the multi-modal encoding layer and the decision control layer based at least on the predicted automatic driving strategy information and the real automatic driving strategy information.

[0120] Since the multi-modal encoding layer and the decision control layer of the model to be trained are connected to form an end-to-end neural network model, the perception information in the sample information can be directly responsible for the decision, and the coupling problem between the prediction and the planning of the trained automatic driving model can be solved. In addition, the introduction of the implicit representation can overcome the defect of the representation of the structured information and cause the algorithm to fail easily. In addition, since the perception information in the sample information can be directly responsible for the decision, the perception can capture information that is critical to the decision, reducing the error accumulation caused by the perception error in the trained model. By introducing the sample interaction information outside the sample vehicle during training, the automatic driving model can determine the automatic driving strategy information based on the surrounding environment and the content of the interaction information, so that the driving behavior of the vehicle can adapt to the requirements of the personnel outside the vehicle.

[0121] In an example, the sample input information can be collected in the process of driving an actual vehicle, for example, collected by a manually driven vehicle with an automatic driving sensor in a real road scene, and the real automatic driving strategy information can be the driving trajectory data of the vehicle in the process of driving in the real road scene (including the control signal recorded for the vehicle in the driving process). In addition, in an example, the sample input information can include sample data collected by an actual vehicle in the process of driving in a real road scene and sample data obtained by a simulated vehicle in the process of driving in a simulated road scene. In some implementations, the sample input information can also include a time identifier of the sample interaction information, a type identifier of the sample interaction information, a time identifier of the sample implicit representation, and a type identifier of the sample implicit representation.

[0122] The sample interaction information can include at least one of action interaction information and language interaction information of the personnel outside the target vehicle. The action interaction information can include a command gesture of the external personnel, such as stop, continue driving, go straight, turn left, turn right, etc. The action interaction information can also include a gesture of the external personnel, such as raising a hand to indicate that the vehicle needs to yield, shaking a hand to indicate that the vehicle does not need to yield, etc. The language interaction information can include the language of the external personnel, for example, the driving behavior of the vehicle can be commanded through shouting.

[0123] In some embodiments, the parameters of the multi-modal encoding layer and the decision control layer can be adjusted based on the objective function in the following formula (1):

[0124]

[0125] where D represents a function for determining the difference between two variables, such as a mean square error function, a cross-entropy function, a KL divergence function, etc. t y represents the predicted automatic driving strategy information output by the automatic driving model at time t, y represents the real driving strategy information at time t.

[0126] After the automatic driving model is trained in the manner based on supervised learning, the automatic driving model can be deployed on a real vehicle, and the parameters of the multi-modal encoding layer and the decision control layer can be adjusted in the process of driving the real vehicle by using the manner of reinforcement learning. For example, reinforcement learning can be performed based on the predicted automatic driving policy information (y1,…, y t ), the real automatic driving policy information , and the sample evaluation feedback information (r1,…, r t ). The sample evaluation feedback information is an evaluation result of the current driving policy information by using artificial evaluation or an evaluation model.

[0127] In an example, the PPO algorithm or the SAC algorithm can be used to perform the reinforcement learning.

[0128] In an example, the objective function in equation (2) can be used to adjust the parameters of the multi-modal encoding layer and the decision control layer:

[0129]

[0130] wherein A t may indicate an advantage function at time t, and A t may be obtained based on the sample evaluation feedback information (r1,…, r t ). α can be a hyperparameter for adjusting the loss value.

[0131] In some embodiments, the decision control layer can be configured to generate the predicted automatic driving policy information based on the sample input information, and generate the predicted interaction information based on the predicted automatic driving policy information.

[0132] In some embodiments, the real automatic driving policy information can include real artificial driving policy information, and the real interaction information can include real artificial interaction information. The real artificial driving policy information (such as vehicle trajectory, control data ) and the real artificial interaction information between the driver and the external person can be obtained by collecting real data of human driving. The real artificial interaction information can include real interaction feedback labels and real natural language content. At the moment when the driver’s conversation occurs, the value of the interaction feedback label is determined to be true (such as 1), and the value of the interaction feedback label at other moments is determined to be false (such as 0).

[0133] Step S440 can include adjusting the parameters of the automatic driving model by using the manner of supervised learning based on the first difference between the predicted automatic driving policy information and the real artificial driving policy information, and the second difference between the predicted interaction information and the real artificial interaction information.

[0134] The objective function for supervised learning can be determined based on equation (3):

[0135]

[0136] wherein L2 is the objective function of supervised learning. D represents a function for determining the difference between two variables, such as a mean square error function, a cross-entropy function, a KL divergence function, etc. t represents the predicted autonomous driving strategy information output by the autonomous driving model at time t, represents the real human driving strategy information at time t, t represents the interaction feedback label output by the autonomous driving model at time t, represents the value of the interaction feedback label in the real data at time t, t represents the predicted natural language content in the predicted interaction information output by the autonomous driving model at time t, represents the real natural language content in the real human interaction information at time t.

[0137] In some embodiments, the real autonomous driving strategy information can include real model driving strategy information, and the real interaction information can include real model interaction information.

[0138] When reinforcement learning training is performed on a real vehicle, the autonomous driving model may need to predict some wrong or failed results, and even may need the target vehicle to collide with surrounding obstacles to learn based on the wrong or collision experience. However, based on the consideration of cost and safety, the autonomous driving vehicle cannot be allowed to have a real collision when training on a real vehicle.

[0139] According to some embodiments, the sample input information can include an intervention label, which can represent whether the real autonomous driving strategy information is autonomous driving strategy information with human intervention.

[0140] Step S440 can include: based on the sample evaluation feedback information (r1,..., r t ), the intervention label (i1,..., i T ), the predicted autonomous driving strategy information (y1,..., y t ), and the real autonomous driving strategy information adjusting the parameters of the multi-modal encoding layer and the decision control layer.

[0141] During the real vehicle training process, the safety officer can intervene at any time in a critical moment to take control of the autonomous vehicle. After the crisis is over, the control is returned to the autonomous vehicle. The intervention identifier is used to represent whether the first real autonomous driving strategy information is the autonomous driving strategy information with human intervention. In other words, by introducing the intervention identifier, the unacceptable model training cost caused by the collision that may occur during real vehicle training can be avoided. Reinforcement learning can gradually learn to avoid adverse situations caused by intervention. Through this mechanism, on the one hand, the efficiency of reinforcement learning can be improved, and on the other hand, the influence of disadvantaged experience on the learning process can be reduced, thereby further improving the robustness of the model obtained by training.

[0142] In an example, the parameters of the multi-modal encoding layer and the decision control layer can be adjusted in a way of feedback reinforcement learning and human-in-the-loop learning. For example, learning can be performed based on five-tuple data including sample evaluation feedback information (r1,..., r t ), intervention identifier (i1,..., i t ), predicted autonomous driving strategy information (y1,..., y t ), real autonomous driving strategy information , and first sample input information (x1,..., x t ).

[0143] Wherein, when the intervention identifier (i1,..., i T ) is a true value, it means that the autonomous vehicle is manually controlled and no longer controlled by the control signal issued by the autonomous driving model; when the intervention identifier (i1,..., i T ) is a non-true value, it means that the autonomous vehicle is controlled by the control signal issued by the autonomous driving model and is not manually controlled.

[0144] In an example, the objective function in equation (4) can be used to adjust the parameters of the multi-modal encoding layer and the evaluation feedback layer as follows:

[0145]

[0146] Wherein, λ1 and λ2 can be hyperparameters indicating the weight of the corresponding component, respectively. Wherein, the intervention identifier (i1,..., i T ) is a true value 1, and a non-true value is 0.

[0147] ​​In a case that the automatic driving model is also capable of outputting predicted interaction information, the data collected in the model on-board test phase can include real model driving strategy information output by the model and real model interaction information. First evaluation feedback information for the real model driving strategy information and second evaluation feedback information for the real model interaction information can be obtained. The first evaluation feedback information and the second evaluation feedback information can be obtained by means of artificial evaluation or by means of a trained evaluation model.

[0148] Step S440 can include adjusting parameters of the automatic driving model based on the first evaluation feedback information, the second evaluation feedback information, a third difference between the predicted automatic driving strategy information and the real model driving strategy information, and a fourth difference between the predicted interaction information and the real model interaction information by means of a reinforcement learning method.

[0149] The objective function for supervised learning can be determined based on equation (5):

[0150]

[0151] wherein L2 is the objective function of the reinforcement learning. D represents a function for determining the difference between two variables, such as a mean square error function, a cross-entropy function, a KL divergence function, etc. t represents predicted automatic driving strategy information output by the automatic driving model at time t, represents real model driving strategy information at time t, t represents predicted natural language content in predicted interaction information output by the automatic driving model at time t, represents real natural language content in real model interaction information at time t. represents an advantage function calculated based on the first evaluation feedback information at time t, represents an advantage function calculated based on the second evaluation feedback information at time t.

[0152] Figure 5 A structural block diagram of an automatic driving device 500 based on an automatic driving model according to an embodiment of the present disclosure is shown. As shown in Figure 5 The device 500 includes:

[0153] An input information obtaining unit 510 is configured to obtain first input information of a multi-modal encoding layer, the first input information including navigation information of a target vehicle and perception information of a surrounding environment of the target vehicle obtained by means of a sensor, the perception information including current perception information and historical perception information of the surrounding environment of the target vehicle in a process of driving the vehicle;

[0154] The multi-modal encoding unit 520 is configured to input the first input information into the multi-modal encoding layer to obtain an implicit representation corresponding to the first input information output by the multi-modal encoding layer; and

[0155] The decision control unit 530 is configured to input second input information including the implicit representation into the decision control layer to obtain target automatic driving strategy information output by the decision control layer, wherein the second input information includes interaction information of a person outside the target vehicle.

[0156] Figure 6 A structural block diagram of a training apparatus 600 of an automatic driving model according to an embodiment of the present disclosure is shown. As shown in the figure, Figure 6 the apparatus 600 includes:

[0157] The sample information obtaining unit 610 is configured to obtain sample input information and real automatic driving strategy information corresponding to the sample input information, the sample input information including sample navigation information of a sample vehicle and sample perception information for a surrounding environment of the sample vehicle, the sample perception information including current sample perception information and historical sample perception information for the surrounding environment of the sample vehicle;

[0158] The multi-modal encoding layer training unit 620 is configured to input the sample input information into the multi-modal encoding layer to obtain a sample implicit representation output by the multi-modal encoding layer;

[0159] The decision control layer training unit 630 is configured to input intermediate sample input information including the sample implicit representation into the decision control layer to obtain predicted automatic driving strategy information output by the decision control layer, wherein the intermediate sample input information includes sample interaction information of a person outside the sample vehicle; and

[0160] The parameter adjusting unit 640 is configured to adjust parameters of the multi-modal encoding layer and the decision control layer based at least on the predicted automatic driving strategy information and the real automatic driving strategy information.

[0161] It should be understood that Figure 5 the various modules or units of the apparatus 500 shown in FIG. 5 can correspond to the various steps in the method 300 described with reference to Figure 3 The operations, features and advantages described above for the method 300 apply equally to the apparatus 500 and the modules and units included therein; and Figure 6 the various modules or units of the apparatus 600 shown in FIG. 6 can correspond to the various steps in the method 300 described with reference to Figure 4The various steps in the described method 400 correspond. Thus, the operations, features and advantages described above for the method 400 apply equally to the apparatus 600 and its comprised modules and units. For the sake of brevity, certain operations, features and advantages are not described again here.

[0162] Although specific functions are discussed above with reference to specific modules, it should be noted that the functions of the various units discussed herein can be split into multiple units and / or at least some of the functions of multiple units can be combined into a single unit.

[0163] It should also be understood that various techniques described herein can be described in the general context of software hardware elements or program modules. The above Figure 5 and Figure 6 The various units described above can be implemented in hardware or in hardware combined with software and / or firmware. For example, the units can be implemented as computer program code / instructions configured to be executed in one or more processors and stored in a computer-readable storage medium. Alternatively, the units can be implemented as hardware logic / circuitry. For example, in some embodiments, one or more of the units 510-530 and the units 610-640 can be implemented together in a System on Chip (SoC). The SoC can include an integrated circuit chip (which includes one or more of a processor (e.g., a Central Processing Unit (CPU), a microcontroller, a microprocessor, a Digital Signal Processor (DSP), etc.), a memory, one or more communication interfaces, and / or other circuitry), and can optionally execute received program code and / or include embedded firmware to perform functions.

[0164] According to another aspect of the present disclosure, an electronic device is also provided, which includes at least one processor, and a memory connected with the at least one processor in communication; the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the automatic driving method or the training method of the automatic driving model according to the embodiments of the present disclosure.

[0165] According to another aspect of the present disclosure, a non-transitory computer-readable storage medium storing computer instructions is also provided, and the computer instructions are used to enable the computer to perform the automatic driving method or the training method of the automatic driving model according to the embodiments of the present disclosure.

[0166] According to another aspect of the present disclosure, a computer program product is also provided, including a computer program, wherein the computer program implements the automatic driving method or the training method of the automatic driving model according to the embodiments of the present disclosure when executed by a processor.

[0167] According to another aspect of the present disclosure, an automatic driving vehicle is also provided, including the automatic driving device 500, the training device 600 of the automatic driving model, and one of the electronic devices described above according to the embodiments of the present disclosure.

[0168] Reference Figure 7 A block diagram of an electronic device 700, which can be an example of a hardware device that can be used in aspects of the present disclosure, will now be described. The electronic device is intended to represent various forms of digital electronic computing devices such as laptops, desktops, workstations, personal digital assistants, servers, blade servers, mainframes, and other appropriate computing devices. The electronic device can also represent various forms of mobile devices such as personal digital processors, cellular telephones, smart phones, wearable devices, and other similar computing devices. The components shown here, their connections and relationships, and their functions, are meant to be examples only, and are not meant to limit implementations of the present disclosure described and / or claimed in this document.

[0169] As Figure 7 shown, the electronic device 700 includes a computing unit 701 that can perform various appropriate actions and processes according to a computer program stored in a read-only memory (ROM) 702 or a computer program loaded into a random access memory (RAM) 703 from a storage unit 708. Various programs and data required for the operation of the electronic device 700 can also be stored in the RAM 703. The computing unit 701, the ROM 702, and the RAM 703 are connected to each other through a bus 704. An input / output (I / O) interface 1705 is also connected to the bus 704.

[0170] A plurality of components in the electronic device 700 are connected to the I / O interface 705, including: an input unit 706, an output unit 707, a storage unit 708, and a communication unit 709. The input unit 706 can be any type of device that can input information to the electronic device 700, and can receive inputted digital or character information, and generate key signal inputs related to user settings and / or function controls of the electronic device, and can include, but is not limited to, a mouse, a keyboard, a touch screen, a track pad, a track ball, a joystick, a microphone, and / or a remote controller. The output unit 707 can be any type of device that can present information, and can include, but is not limited to, a display, a speaker, a video / audio output terminal, a vibrator, and / or a printer. The storage unit 708 can include, but is not limited to, a magnetic disk, an optical disk. The communication unit 709 allows the electronic device 700 to exchange information / data with other devices through a computer network such as the Internet and / or various telecommunication networks, and can include, but is not limited to, a modem, a network card, an infrared communication device, a wireless communication transceiver, and / or a chipset, such as a Bluetooth device, an 802.11 device, a WiFi device, a WiMax device, a cellular communication device, and / or the like.

[0171] The computing unit 701 can be various general and / or special purpose processing components having processing and computing capabilities. Some examples of the computing unit 701 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, a digital signal processor (DSP), and any appropriate processor, controller, microcontroller, etc. The computing unit 701 performs various methods and processes described above, such as the methods (or processes) 300 to 400. For example, in some embodiments, the methods (or processes) 300 to 400 can be implemented as a computer software program tangibly embodied in a machine-readable medium, such as the storage unit 708. In some embodiments, part or all of the computer program can be loaded and / or installed onto the electronic device 700 via the ROM 702 and / or the communication unit 709. When the computer program is loaded onto the RAM 703 and executed by the computing unit 701, one or more steps of the methods (or processes) 300 to 400 described above can be performed. Alternatively, in other embodiments, the computing unit 701 can be configured to perform the methods (or processes) 300 to 400 by any other appropriate means, such as by means of firmware.

[0172] The various embodiments of the systems and techniques described above can be implemented in digital electronic circuitry, integrated circuitry, a field programmable gate array (FPGA), an application specific integrated circuit (ASIC), a system on a chip (SOC), a load programmable logic device (CPLD), computer hardware, firmware, software, and / or combinations thereof. These various embodiments can include implementation in one or more computer programs that are executable and / or interpretable on a programmable system including at least one programmable processor, which can be special or general purpose, coupled to receive data and instructions from, and to transmit data and instructions to, a storage system, at least one input device, and at least one output device.

[0173] Program code for carrying out methods of the present disclosure can be written in any combination of one or more programming languages. The program code can be provided to a processor or controller of a general purpose computer, special purpose computer, or other programmable data processing apparatus to produce a machine, such that the program code, when executed by the processor or controller, produces the functions / operations specified in the flowcharts and / or the block diagrams. The program code can be executed entirely on a machine, partially on a machine, partially on a machine as a stand-alone software package, or entirely on a remote machine or server.

[0174] In the context of the present disclosure, a machine-readable medium can be a tangible medium that contains or stores a program for use by or in connection with an instruction execution system, apparatus, or device. The machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium can include but is not limited to an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any suitable combination of the foregoing. More specific examples of the machine-readable storage medium will include one or more lines of electrical conductors, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or Flash memory), an optical fiber, a portable compact disc read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing.

[0175] To provide for interaction with a user, the systems and techniques described here can be implemented on a computer having a display device (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor) for displaying information to the user and a keyboard and a pointing device (e.g., a mouse or a trackball) by which the user can provide input to the computer. Other kinds of devices can be used to provide for interaction with a user as well; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form, including acoustic, speech, or tactile input.

[0176] The systems and techniques described here can be implemented in a computing system that includes a back end component (e.g., as a data server), or that includes a middleware component (e.g., an application server), or that includes a front end component (e.g., a user computer having a graphical user interface or a Web browser through which a user can interact with an implementation of the systems and techniques described here), or any combination of such back end, middleware, or front end components. The components of the system can be interconnected by any form or medium of digital data communication (e.g., a communication network). Examples of communication networks include a local area network (LAN), a wide area network (WAN), the Internet, and a blockchain network.

[0177] The computer system can include clients and servers. A client and server are generally remote from each other and typically interact through a communication network. The relationship of client and server is generally established by computer programs running on the respective computers and having a client-server relationship to each other. The servers can be cloud servers, servers of a distributed system, or servers combined with a blockchain.

[0178] It should be understood that the various forms of flow shown above can be re-ordered, added to, or deleted from without departing from the scope of the present disclosure. For example, the steps recited in the present disclosure can be performed in parallel, sequentially, or in a different order, as long as the desired results of the technology disclosed in the present disclosure are achieved, which is not limited herein.

[0179] While embodiments or examples of this disclosure have been described with reference to the figures, it will be understood that the methods, systems, and devices described above are merely exemplary embodiments or examples, and the scope of the application is not limited to these embodiments or examples. Various elements of the embodiments or examples can be omitted or substituted by equivalents thereof. Furthermore, the steps can be performed in a different order than described in the disclosure. Further, various elements of the embodiments or examples can be combined in various ways. It is important that as technology evolves, many of the elements described herein can be substituted by equivalents which serve the same function.

Claims

1. An autonomous driving model system, comprising a multi-modal encoding layer and a decision control layer, the multi-modal encoding layer and the decision control layer being connected to form an end-to-end neural network model, such that the decision control layer directly obtains autonomous driving policy information based on an output of the multi-modal encoding layer, wherein the multi-modal encoding layer is configured to encode first input information to obtain an implicit representation corresponding to the first input information, the first input information comprising navigation information of a target vehicle, perception information of a surrounding environment of the target vehicle obtained by a sensor, and interaction information with a person outside the target vehicle, the perception information comprising current perception information and historical perception information of the surrounding environment of the target vehicle during driving of the target vehicle, and the decision control layer is configured to obtain target autonomous driving policy information based on second input information comprising the implicit representation. wherein 2.The autonomous driving model system of claim 1, wherein the interaction information from outside the target vehicle comprises at least one of action interaction information and language interaction information. The interaction information in the second input information is included in the implicit representation. 4.The autonomous driving model system of claim 2, wherein the second input information further comprises a time identifier of the interaction information, a type identifier of the interaction information, a time identifier of the implicit representation, and a type identifier of the implicit representation.

3. The automated driving model system of claim 2, wherein, 5.The autonomous driving model system of claim 1, wherein the decision control layer is further configured to generate target interaction information based on the second input information. 6.The autonomous driving model system of claim 5, wherein the target interaction information comprises an interaction feedback identifier, and when the interaction feedback identifier is true, the target interaction information further comprises interaction content. 7.An autonomous driving method implemented by an autonomous driving model, the autonomous driving model comprising a multi-modal encoding layer and a decision control layer, the multi-modal encoding layer and the decision control layer being connected to form an end-to-end neural network model, such that the decision control layer directly obtains autonomous driving policy information based on an output of the multi-modal encoding layer, the method comprising: obtaining first input information of the multi-modal encoding layer, the first input information comprising navigation information of a target vehicle, perception information of a surrounding environment of the target vehicle obtained by a sensor, and interaction information with a person outside the target vehicle, the perception information comprising current perception information and historical perception information of the surrounding environment of the target vehicle during driving of the target vehicle; inputting the first input information into the multi-modal encoding layer to obtain an implicit representation corresponding to the first input information, the implicit representation being obtained by encoding the first input information by the multi-modal encoding layer; and inputting second input information comprising the implicit representation into the decision control layer to obtain target autonomous driving policy information output by the decision control layer. ​ ​ ​ ​ ​ ​ 8.The automatic driving method of claim 7, wherein the interaction information from outside the target vehicle comprises at least one of action interaction information and language interaction information.

9. The automatic driving method of claim 8, wherein, The action interaction information in the second input information is included in the implicit representation. 10.The automatic driving method of claim 8, wherein the second input information further comprises a time identifier of the language interaction information, a type identifier of the language interaction information, a time identifier of the implicit representation, and a type identifier of the implicit representation. 11.The automatic driving method of claim 7, wherein the decision control layer is further configured to generate target interaction information based on the second input information. 12.The automatic driving method of claim 11, wherein the target interaction information comprises an interaction feedback identifier, and when the interaction feedback identifier is true, the target interaction information further comprises interaction content. 13.A method for training an automatic driving model, the automatic driving model comprising a multi-modal encoding layer and a decision control layer, the multi-modal encoding layer and the decision control layer being connected to form an end-to-end neural network base model, such that the decision control layer directly obtains automatic driving policy information based on an output of the multi-modal encoding layer, the method comprising: obtaining sample input information and real automatic driving policy information corresponding to the sample input information, the sample input information comprising sample navigation information of a sample vehicle, sample perception information for a surrounding environment of the sample vehicle, and sample interaction information with a person outside the sample vehicle, the sample perception information comprising current sample perception information and historical sample perception information for the surrounding environment of the sample vehicle; inputting the sample input information into the multi-modal encoding layer to obtain a sample implicit representation obtained by encoding the sample input information by the multi-modal encoding layer; inputting intermediate sample input information comprising the sample implicit representation into the decision control layer to obtain predicted automatic driving policy information output by the decision control layer; and adjusting parameters of the multi-modal encoding layer and the decision control layer based at least on the predicted automatic driving policy information and the real automatic driving policy information. 14.The method of claim 13, wherein the sample input information further comprises a time identifier of the sample interaction information, a type identifier of the sample interaction information, a time identifier of the sample implicit representation, and a type identifier of the sample implicit representation. The decision control layer is configured to generate predicted automatic driving policy information based on the sample input information, and to generate predicted interaction information based on the predicted automatic driving policy information. The real automatic driving policy information comprises real human driving policy information, and the real interaction information comprises real human interaction information.

15. The method of claim 13, wherein, adjusting parameters of the multi-modal encoding layer and the decision control layer based at least on the predicted automatic driving policy information and the real automatic driving policy information comprises:

16. The method of claim 15, wherein, ​ ​ The parameters of the automatic driving model are adjusted based on a first difference between the predicted automatic driving strategy information and the real manual driving strategy information and a second difference between the predicted interaction information and the real manual interaction information in a manner of supervised learning. 17.The method of claim 16, wherein the real automatic driving strategy information comprises real model driving strategy information, and the real interaction information comprises real model interaction information, and the method further comprises: obtaining first evaluation feedback information for the real model driving strategy information and second evaluation feedback information for the real model interaction information; The parameters of the automatic driving model are adjusted based on a third difference between the first evaluation feedback information, the second evaluation feedback information, the predicted automatic driving strategy information and the real model driving strategy information and a fourth difference between the predicted interaction information and the real model interaction information in a manner of reinforcement learning. 18.An automatic driving device based on an automatic driving model, the automatic driving model comprising a multi-modal encoding layer and a decision control layer, the multi-modal encoding layer and the decision control layer being connected to form an end-to-end neural network model, so that the decision control layer directly obtains automatic driving strategy information based on an output of the multi-modal encoding layer, the device comprising: an input information obtaining unit configured to obtain first input information of the multi-modal encoding layer, the first input information comprising navigation information of a target vehicle and perception information of a surrounding environment of the target vehicle obtained by a sensor and interaction information of a person outside the target vehicle, the perception information comprising current perception information and historical perception information of the surrounding environment of the target vehicle during driving of the vehicle; a multi-modal encoding unit configured to input the first input information into the multi-modal encoding layer to obtain an implicit representation corresponding to the first input information obtained by encoding the first input information by the multi-modal encoding layer; and a decision control unit configured to input second input information comprising the implicit representation into the decision control layer to obtain target automatic driving strategy information output by the decision control layer. 19.A training device of an automatic driving model, the automatic driving model comprising a multi-modal encoding layer and a decision control layer, the multi-modal encoding layer and the decision control layer being connected to form an end-to-end neural network model, so that the decision control layer directly obtains automatic driving strategy information based on an output of the multi-modal encoding layer, the device being configured to train the multi-modal encoding layer and the decision control layer, and comprising: a sample information obtaining unit configured to obtain sample input information and real automatic driving strategy information corresponding to the sample input information, the sample input information comprising sample navigation information of a sample vehicle and sample perception information of a surrounding environment of the sample vehicle and sample interaction information of a person outside the sample vehicle, the sample perception information comprising current sample perception information and historical sample perception information of the surrounding environment of the sample vehicle; a multi-modal encoding layer training unit, configured to input the sample input information into the multi-modal encoding layer to obtain a sample implicit representation obtained by encoding the sample input information by the multi-modal encoding layer; a decision control layer training unit, configured to input intermediate sample input information comprising the sample implicit representation into the decision control layer to obtain predicted autonomous driving strategy information output by the decision control layer; and a parameter adjustment unit, configured to adjust parameters of the multi-modal encoding layer and the decision control layer based at least on the predicted autonomous driving strategy information and real autonomous driving strategy information.

20. An electronic device, comprising: at least one processor; and a memory connected to the at least one processor in communication; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the method of any one of claims 7-17. The computer instructions are used to enable the computer to perform the method according to any one of claims 7-17.

21. A non-transitory computer readable storage medium having stored thereon computer instructions, wherein, The computer program, when executed by a processor, implements the method of any one of claims 7-17.

22. A computer program product comprising a computer program, wherein, 23. An autonomous driving vehicle, comprising: one of the autonomous driving apparatus of claim 18, the training apparatus of the autonomous driving model of claim 19, and the electronic device of claim 20. ​

Images