Controllable and explainable automatic driving method and system

Abstract

According to the controllable and explainable automatic driving method and system, an intelligent vehicle generates strategies from historical driving behaviors and environment feedback in an induction mode, and decisions can be controlled through the strategies. The system comprises a cognitive model, a decision model and a value model. The cognitive model receives the input of the environment, determines which strategy is used, and can carry out strategy induction through the decision-making effect or historical driving behaviors. And the decision model receives the input of the environment and the strategy output by the cognitive model and makes a decision, execution is performed after decision making, and learning is performed through environment feedback. The value model is used for judging which decision of the current driving state is good so as to help the decision model to learn. Therefore, the system can automatically generate a cognitive driving strategy and control a decision through the driving strategy.

Description

Technical Field

[0001] The present invention relates to the field of autonomous driving technology, and in particular to a controllable and explainable autonomous driving method and system. Background Art

[0002] Traditional autonomous driving systems usually rely on predefined rules and limited sensor data for environmental perception and driving decision-making, which makes them limited when facing complex and dynamically changing driving environments, especially when dealing with extreme weather, emergencies or ambiguous traffic rules. The decision-making effect may not be ideal. In addition, many current autonomous driving systems still have challenges in the controllability and explainability of the decision-making process and the generation of optimization strategies, and cannot fully meet the needs of highly autonomous driving. Summary of the invention

[0003] In order to overcome the shortcomings of the above-mentioned prior art, the purpose of the present invention is to provide a controllable and explainable autonomous driving method and system, so that artificial intelligence can self-generate cognitive strategies and be controllable in the decision-making process, and further enable the decision-making of artificial intelligence to understand and adapt to the environment and be subject to external management.

[0004] In order to achieve the above object, the technical solution adopted by the present invention is:

[0005] The present invention provides a controllable and explainable autonomous driving method and system, comprising the following steps:

[0006] Step 1: Establish cognitive model, decision model and value model;

[0007] Step 2: During the operation phase, the cognitive model receives driving environment information and selects a strategy, and the decision model uses the driving environment information and the strategy as input to make a decision; wherein the driving environment information refers to information that can be used for decision-making in an autonomous driving scenario; the strategy is a collection of a series of driving operations of the vehicle; and the decision refers to deciding which driving operation to perform;

[0008] Step 3: By fixing the strategy selected by the cognitive model, the decision of the decision model is controlled to achieve controllable decision-making of the intelligent agent;

[0009] Step 4: In the self-learning training stage, the cognitive model receives driving environment information and selects a strategy, and the decision model uses the driving environment information and the strategy as input to make a decision; after performing a corresponding driving operation according to the decision, the value model collects the driving results and receives the driving environment information as input to evaluate the pros and cons of the corresponding driving operation;

[0010] Step 5: The decision model is optimized according to the evaluation results of the value model, the value model is optimized according to the gap between the environment and the expectation after the driving operation is performed, and the cognitive model generates a strategy based on the historical driving results;

[0011] Step 6: During the imitation learning training phase, appropriate driving data is collected, such as driving environment information, driving operation sequences, and driving results during driving of excellent driving cases. The cognitive model receives the driving environment information during driving and selects a strategy. The decision model uses the driving environment information and the strategy as input to make a decision. The decision model is optimized based on whether the decision matches the collected driving operation sequence. The cognitive model generates a strategy based on the driving results.

[0012] In one embodiment, the step 1 specifically includes the following steps:

[0013] Step 1.1: Establish the cognitive model using an artificial neural network and an embedding space containing a strategy representation vector, wherein the input of the cognitive model is driving environment information, and the strategy representation vector is selected to form an output; the strategy representation vector is a vector with a length of d;

[0014] Step 1.2: Using an artificial neural network to establish the decision model, the input of the decision model is the driving environment information and its corresponding strategy representation vector, and the output is the driving operation to be performed;

[0015] Step 1.3: Use an artificial neural network to establish the value model, the input of the value model is the driving environment information, and the output is an evaluation value of whether the driving operation to be performed in the current decision is suitable for the current driving environment.

[0016] In one embodiment, the size of the embedding space is n*d, where n represents the number of strategy representation vectors contained therein, which can be expanded with learning, and d is the length of the strategy representation vector. All strategy representation vectors are initially randomized; the driving environment information is input into the artificial neural network of the cognitive model, and a vector with a length of d is output; the similarity between the vector with a length of d and the strategy representation vector in the embedding space is calculated, and a suitable strategy vector in the embedding space is selected for output, or a combination of multiple strategy representation vectors in the embedding space can be selected for output.

[0017] In one embodiment, step 2 specifically includes the following steps:

[0018] Step 2.1: Use artificial neural network to establish feature extraction sub-models for both cognitive model and decision model. These two feature extraction sub-models can be the same feature extraction sub-model, that is, these two feature extraction sub-models can share weights;

[0019] Step 2.2: Input the driving environment information into the feature extraction sub-model, and the feature extraction sub-model outputs the corresponding environment representation vector;

[0020] Step 2.3: Input the corresponding environment representation vector into the cognitive model, and the cognitive model outputs the strategy representation vector;

[0021] Step 2.4: The decision model receives the corresponding environment representation vector and strategy representation vector, selects the optimal driving operation judged by the decision model as output, and executes it.

[0022] In one embodiment, step 3 specifically includes the following steps:

[0023] Step 3.1: During runtime, the decision is controlled. At this time, the cognitive model does not accept driving environment information as input, and directly selects a specific strategy representation vector in the embedding space as output. The specific strategy representation vector is selected according to the corresponding control effect after the strategy interpretation;

[0024] Step 3.2: The decision model receives the corresponding environment representation vector and strategy representation vector, selects the optimal driving operation judged by the decision model as output, and executes it.

[0025] In one embodiment, step 4 specifically includes the following steps:

[0026] Step 4.1: According to step 2.1, use artificial neural network to establish the feature extraction sub-model of the value model;

[0027] Step 4.2: Input the driving environment information into the feature extraction sub-model, and the feature extraction sub-model outputs the corresponding environment representation vector;

[0028] Step 4.3: Input the corresponding environment representation vector into the cognitive model, and the cognitive model outputs the strategy representation vector. The decision model receives the corresponding environment representation vector and strategy representation vector, selects the optimal driving operation judged by the decision model as output, and executes it;

[0029] Step 4.4: The value model accepts the corresponding environment representation vector and outputs the gap between the environment and the expectation after the driving operation is performed.

[0030] In one embodiment, step 5 specifically includes the following steps:

[0031] Step 5.1: The decision model is optimized using gradient backpropagation according to the evaluation results of the value model as follows:

[0032] Step 5.1.1: Combine the decision model output of the selected optimal driving action described in step 4.3 and the gap described in step 4.4 into a gradient;

[0033] Step 5.1.2: Optimize the various components of the gradient backpropagation decision model described in step 5.1.1;

[0034] Step 5.2: The value model is optimized using an error back propagation algorithm based on the gap between the environment and the expectation after the driving operation is performed. The process is as follows:

[0035] Step 5.2.1: The gap between the environment and the expectation after the driving operation is performed is used as the loss function;

[0036] Step 5.2.2: Use the loss function described in step 5.2.1 to optimize the error back propagation algorithm;

[0037] Step 5.3: The cognitive model generates a strategy based on historical driving results. The process is as follows:

[0038] Step 5.3.1: The cognitive model receives and records the driving results;

[0039] Step 5.3.2: The optimization objectives of the cognitive model include two: the first is that the output vector of the artificial neural network of the cognitive model and the selected strategy representation vector have as large a similarity as possible; the second is to use the artificial neural network to construct a reconstruction sub-model, the reconstruction sub-model takes the driving operation executed by historical decision and the driving result after execution, the driving environment information during execution and the selected strategy representation vector as input, and the output is the optimal driving operation, and the matching of the output driving operation and the executed driving operation is used as the loss function;

[0040] Step 5.3.3: Use the error back-propagation algorithm to update the artificial neural network and embedding space of the cognitive model.

[0041] In one embodiment, step 6 specifically includes the following steps:

[0042] Step 6.1: During the imitation learning training phase, appropriate driving data is collected, such as driving environment information, driving operation sequences, and driving results during driving;

[0043] Step 6.2: The cognitive model receives the driving environment information during driving and then selects a strategy as described in step 2, and the decision model takes the driving environment information and the strategy as input and makes a decision as described in step 2;

[0044] Step 6.3: The decision model is optimized as described in step 5 based on whether the decision and the collected driving operation sequence match as the error function;

[0045] Step 6.4: The cognitive model generates strategies by summarizing the driving results according to the method described in step 5.

[0046] The second aspect of the present invention provides a controllable intelligent agent decision system based on a self-generated cognitive strategy, the interpretable method of which is:

[0047] Step 7.1: Collect historical driving decision-making processes and determine the physical meaning or behavior pattern of the strategy represented by each decision representation vector in the embedding space;

[0048] Step 7.2: Explain the strategy chosen during driving.

[0049] The third invention of the present invention provides a controllable and explainable autonomous driving system, which includes a processor, and the processor is used to execute a controllable and explainable autonomous driving method described in the present invention.

[0050] Compared with the prior art, the present invention has the following beneficial effects:

[0051] 1. Controllability of the decision-making process: The present invention enables the autonomous driving system to control the decision-making process according to predefined strategies by introducing the collaborative work of cognitive models, decision models and value models. By fixing the strategy selected by the cognitive model, the output of the decision model can be effectively controlled to achieve the controllability of the decision-making process of the autonomous driving system. This can provide higher flexibility and safety when dealing with complex driving scenarios or specific needs.

[0052] 2. Self-learning and imitation learning capabilities: The present invention designs a self-learning and imitation learning training mechanism, which enables the system to continuously improve driving strategies and decision-making effects by continuously optimizing cognitive models, decision models and value models. In the self-learning stage, the driving results are evaluated and driving operations are optimized through feedback from the value model, while in the imitation learning stage, the driving strategy of the system is further enhanced through the input of actual driving data, and the accuracy and adaptability of driving decisions are improved.

[0053] 3. Explainability of decision-making: Traditional autonomous driving systems often lack transparency in the decision-making process and it is difficult to explain the reasons for their choices. However, this invention can explain the decisions made during the autonomous driving process by analyzing the physical meaning or behavior pattern of the strategy representation vector in the embedding space. This enhances the transparency and credibility of the system for users and regulators. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 It is a schematic diagram of the system structure of the present invention.

[0055] Figure 2 It is a schematic diagram of the operation process of the present invention.

[0056] Figure 3 It is a schematic diagram of the self-learning implementation of the present invention.

[0057] Figure 4 It is a schematic diagram of the implementation of the strategy of the present invention.

[0058] Figure 5 It is a schematic diagram of the imitation learning implementation of the present invention. DETAILED DESCRIPTION

[0059] The embodiments of the present invention are described in detail below with reference to the accompanying drawings and examples.

[0060] In order to meet the requirements of controllability, explainability and optimization strategy generation of the decision-making process, an autonomous driving system based on artificial neural networks can be considered. By introducing cognitive models, decision models and value models, not only can adaptive environmental perception and decision-making be achieved, but also strategies can be continuously optimized during the self-learning process to improve the overall performance of the system. However, there is currently a lack of an autonomous driving system that can organically combine cognitive models, decision models and value models and has multi-stage training and optimization capabilities to cope with complex driving scenarios.

[0061] To this end, the present invention proposes an autonomous driving method and system based on artificial neural networks that coordinate cognition, decision-making, and value models to achieve controllability and explainability of the decision-making process, and further optimize driving operations and strategies through a multi-stage training process, thereby improving the safety, stability, and adaptability of the system.

[0062] The automatic driving method of the present invention can be specifically described as the following steps:

[0063] Step 1: Establish cognitive model, decision model and value model.

[0064] like Figure 1 As shown, this step can be specifically described as the following process:

[0065] Step 1.1: Use an artificial neural network and an embedding space containing a strategy representation vector to establish the cognitive model. In this example, the artificial neural network is a long short-term memory recurrent neural network (LSTM). The input of the cognitive model is driving environment information. In this example, the driving environment information includes but is not limited to vehicle camera data, vehicle radar data, etc., which can be obtained while the vehicle is driving. The strategy representation vector is selected to form the output. The strategy representation vector is a vector of length d.

[0066] The embedding space can be specifically described as: the size is n*d, n represents the number of strategy representation vectors contained, which can be expanded with learning, d is the length of the strategy representation vector, and all strategy representation vectors are randomized initially; the driving environment information is input into the artificial neural network of the cognitive model, and a vector with a length of d is output; the similarity between the vector with a length of d and the strategy representation vector in the embedding space is calculated, and in this example, multiple strategy representation vectors in the embedding space are selected for combined output.

[0067] In this example, the first k strategy representation vector combinations with the largest similarity are selected, and the combination method is to concatenate these strategy representation vectors to obtain a long vector.

[0068] Step 1.2: Use an artificial neural network to establish the decision model. In this example, a fully connected network is selected. The input of the decision model is the driving environment information and its corresponding strategy representation vector (i.e., the output of the cognitive model), and the output is the driving operation that needs to be performed.

[0069] Step 1.3: Use an artificial neural network to establish the value model. In this example, a fully connected network is selected. The input of the value model is driving environment information, and the output is an evaluation value of whether the driving operation required to be performed in the current decision is suitable for the current driving environment. The larger the evaluation value, the more suitable the current driving operation.

[0070] Step 2: If Figure 2 As shown, in the operation stage, the cognitive model receives driving environment information and then selects a strategy, and the decision model uses the driving environment information and the strategy as input to make a decision; the decision refers to deciding what driving operation to perform, which in this example includes: whether there is information on the steering direction of the steering wheel and the throttle within a unit time of 0.1 seconds.

[0071] This step can be specifically described as the following process:

[0072] Step 2.1: Use artificial neural networks to establish feature extraction sub-models for both cognitive models and decision-making models. In this example, a fully connected artificial neural network is selected, and these two feature extraction sub-models share weights.

[0073] Step 2.2: Input the driving environment information into the feature extraction sub-models of the cognitive model and the decision model, and each feature extraction sub-model outputs its corresponding environment representation vector.

[0074] Step 2.3: Input the corresponding environment representation vector into the cognitive model, the cognitive model performs similarity calculation and outputs the strategy representation vector.

[0075] Step 2.4: The decision model receives the corresponding environment representation vector and strategy representation vector, selects the optimal driving operation judged by the decision model as output, and executes it, that is, the autonomous driving car.

[0076] Step 3: By fixing the strategy selected by the cognitive model, the decision of the decision model is controlled to achieve controllable decision-making of the intelligent agent.

[0077] This step can be specifically described as the following process:

[0078] Step 3.1: At runtime, the decision can be controlled. At this time, the cognitive model does not accept driving environment information as input, and directly selects a specific strategy representation vector as output in the embedding space. The specific selection here can be selected according to the corresponding control effect after interpreting the strategy according to step 7. For example, some strategy representation vectors represent low-speed and slow driving strategies, and these strategies can be selected to control the autonomous driving.

[0079] Step 3.2: The decision model receives the corresponding environment representation vector and strategy representation vector, selects the optimal driving operation judged by the decision model as output, and executes it, that is, the autonomous driving car.

[0080] Step 4: Figure 3 As shown, in the self-learning training stage, the cognitive model receives driving environment information and selects a strategy, and the decision model uses the driving environment information and the strategy as input to make a decision; after performing the corresponding driving operation according to the decision, the value model collects the driving results and receives the driving environment information as input to evaluate the quality of this driving operation.

[0081] This step can be specifically described as the following process:

[0082] Step 4.1: Based on step 2.1, a feature extraction sub-model of the value model is established. In this example, a fully connected artificial neural network is selected.

[0083] Step 4.2: Input the driving environment information into each feature extraction sub-model, and each feature extraction sub-model outputs its corresponding environment representation vector.

[0084] Step 4.3: Input the corresponding environment representation vector into the cognitive model, and the cognitive model outputs the strategy representation vector. The decision model receives the corresponding environment representation vector and the strategy representation vector, and selects the optimal driving operation judged by the decision model as the output. In this example, the optimal method is the driving operation with the largest output value of the decision model, and executes it.

[0085] Step 4.4: The value model receives the corresponding environment representation vector and outputs the gap between the environment and the expectation after the driving operation is performed. In this example, the value model receives the current environment representation vector and the environment representation vector after the driving operation is performed, outputs the value / expected value of the current environment and the environment after the operation is performed, and then receives the reward of the driving environment feedback after the driving operation is performed, measures the gap between the expectation and the reward, and uses this gap, that is, the evaluation result of the value model, to optimize the model in the subsequent steps.

[0086] Step 5: The decision model is optimized according to the evaluation results of the value model, the cognitive model generates strategies according to historical driving results, and the value model is optimized according to the gap between the environment and expectations after the driving operation is performed.

[0087] This step can be specifically described as the following process:

[0088] Step 5.1: The decision model is optimized using gradient backpropagation according to the evaluation results of the value model.

[0089] Step 5.1.1: Combine the decision model output of the selected optimal driving action described in step 4.3 and the gap described in step 4.4 into a gradient.

[0090] Step 5.1.2: Pass the gradient described in step 5.1.1 back to the feature extraction sub-model and other artificial neural network components of the decision model for gradient descent optimization.

[0091] Step 5.2: The value model is optimized using an error back propagation algorithm according to the gap between the environment and the expectation after performing the driving operation.

[0092] Step 5.2.1: The gap between the environment and the expectation after performing the driving operation is used as the loss function.

[0093] Step 5.2.2: Use the loss function described in step 5.2.1 to optimize the error back propagation algorithm.

[0094] Step 5.3: The cognitive model generates a strategy based on historical driving results.

[0095] Step 5.3.1: The cognitive model receives and records the driving results.

[0096] Step 5.3.2: If Figure 4 As shown in , the optimization objectives of the cognitive model are divided into two: the first is that the output vector of the artificial neural network of the cognitive model and the selected strategy representation vector are as similar as possible. In this example, the mean square error loss is used to measure the similarity. When selecting the strategy representation vector, the selection is made based on the similarity, as shown in Figure 4As shown in the lower part, the similarity values ​​between the d-dimensional vector output by the artificial neural network and the vectors representing each strategy in the embedding space are calculated, and the largest k are selected. After selection, gradient backpropagation optimization is performed with the goal of maximizing these similarity values ​​during optimization; the second is to use the artificial neural network to build a reconstruction sub-model, such as Figure 4 As shown in the upper part, the reconstructed sub-model takes the driving operation executed by historical decisions and the driving results after execution, the driving environment information during execution, and the selected strategy representation vector as input, and the output is the optimal driving operation. The output driving operation and the actual driving operation are matched as the loss function. In this example, the cross entropy function loss is used for matching, and this is used as the loss index for gradient backpropagation optimization.

[0097] Step 5.3.3: Use the error back-propagation algorithm to update the artificial neural network and embedding space of the cognitive model.

[0098] Step 6: Figure 5 As shown, in the imitation learning training stage, appropriate driving data is collected, such as driving environment information, driving operation sequence and driving results during driving, the cognitive model receives the driving environment information during driving and selects a strategy, and the decision model uses the driving environment information and the strategy as input to make a decision. The decision model is optimized based on whether the decision and the collected driving operation sequence match, and the cognitive model generates a strategy based on the driving results.

[0099] This step can be specifically described as the following process:

[0100] Step 6.1: During the imitation learning training phase, collect appropriate driving data, such as driving environment information during driving, driving operation sequences (such as Figure 5 actual operations in the game) and driving results, etc.

[0101] Step 6.2: The cognitive model receives the driving environment information during driving and then selects a strategy as described in steps 2.1-2.3. The decision model takes the driving environment information and the strategy as input and makes a decision as described in step 2.4.

[0102] Step 6.3: The decision model uses the error function based on whether the decision matches the driving operation sequence collected in step 6.1. In this example, the match is measured using the cross entropy function and optimized as described in step 5.1.

[0103] Step 6.4: The cognitive model generates strategies by summarizing the driving results according to the method described in step 5.3.

[0104] The steps explained in the present invention can be specifically described as the following process:

[0105] Step 7.1: Collect historical driving decision-making processes and determine the physical meaning or behavior pattern of the strategy represented by each decision representation vector in the embedding space.

[0106] Step 7.2: Explain the strategy selected during driving; the present invention can thereby obtain a controllable intelligent agent decision system based on self-generated strategies. The system mainly includes a decision model and a cognitive model. The cognitive model is used to select a suitable strategy in the current environment according to the received environmental information. The decision model is used to make decisions in combination with the environment and the strategy. After the decision is made, it is executed and learned by environmental feedback. The cognitive model generates a strategy by induction after environmental feedback. The system can thus self-generate cognitive strategies and control decisions through strategies. Compared with other artificial intelligence decisions, the present invention can control the decision-making process by controlling the cognitive model.

[0107] In the CARLA simulated autonomous driving task, the average deviation between the actual running path and the ideal planned path is about 5.7 cm, which is better than the 9.6 cm of SAC and 13.4 cm of PPO of the same type of reinforcement learning algorithms.

[0108] The above CARLA provides the concept of waypoints, and all vehicle information can be obtained through waypoints. Waypoints can be understood as discrete points distributed along the center line of the lane. The minimum distance between these discrete points in CARLA is 2 cm. The process of a vehicle driving on the road is the process of passing through one waypoint after another. The ideal planned path is the set of optimal driving path points from point A to point B under the current traffic scenario. The actual operation path is the set of real driving path points. By comparing the deviation values ​​of the actual operation path and the path points in the ideal planned path at the same time, the average value of the overall deviation is calculated.

[0109] The present invention statistically analyzes the strategies in the embedded space to summarize the different driving environment information and driving decisions corresponding to different strategies. For example, in a driving scenario where the right auxiliary road merges into the main road and then exits from the right auxiliary road, most of the strategies used show an obvious tendency to turn right in the statistical analysis. In the straight driving scenario, the strategy used basically maintains straight driving in the statistical analysis. Therefore, the system of the present invention that executes the above-mentioned automatic driving method optimizes and controls driving decisions during actual driving through the collaborative work of cognitive models, decision models and value models, can autonomously adapt to complex and changeable driving environments, and optimize driving strategies and operations during the training stages of self-learning and imitation learning, thereby improving the robustness and decision transparency of the automatic driving system.

[0110] The embodiment of the present invention further provides a computer storage medium on which a computer program is stored, and when the computer program is executed by a computer, the computer can perform the method of the above method embodiment. In other words, the embodiment of the present invention further provides a computer program product containing instructions, and when the instructions are executed by a computer, the computer can perform the method of the above method embodiment.

[0111] When software is used for implementation, it can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, the process or function according to the embodiment of the present invention is generated in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices. The computer instructions can be stored in a computer-readable storage medium, or transmitted from one computer-readable storage medium to another computer-readable storage medium. For example, the computer instructions can be transmitted from a website site, a computer, a server or a data center to another website site, a computer, a server or a data center by wired or wireless means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or a data center that includes one or more available media integration. The available medium can be a magnetic medium, an optical medium, or a semiconductor medium, etc.

[0112] The method or system of the present invention can be implemented in an electronic device, and the method of the present invention can be executed in the electronic device, and the system obtained by the training of the present invention can also be implanted in the electronic device. Further, the electronic device can include a collection unit and a final operation unit, and the game information collected by the collection unit can be directly or after necessary pre-processing sent to the system of the present invention, and the operation is performed in the operation unit.

Claims

1. A controllable and explainable autonomous driving method, characterized in that: The steps include: Step 1: Establish cognitive model, decision model and value model; Step 2: During the operation phase, the cognitive model receives driving environment information and selects a strategy, and the decision model uses the driving environment information and the strategy as input to make a decision; wherein the driving environment information refers to information that can be used for decision-making in an autonomous driving scenario; the strategy is a collection of a series of driving operations of the vehicle; and the decision refers to deciding which driving operation to perform; Step 3: By fixing the strategy selected by the cognitive model, the decision of the decision model is controlled to achieve controllable decision-making of the intelligent agent; Step 4: In the self-learning training stage, the cognitive model receives driving environment information and selects a strategy, and the decision model uses the driving environment information and the strategy as input to make a decision; after performing a corresponding driving operation according to the decision, the value model collects the driving results and receives the driving environment information as input to evaluate the pros and cons of the corresponding driving operation; Step 5: The decision model is optimized according to the evaluation results of the value model, the value model is optimized according to the gap between the environment and the expectation after the driving operation is performed, and the cognitive model generates a strategy based on the historical driving results; Step 6: In the imitation learning training stage, the cognitive model receives the driving environment information during driving and selects a strategy. The decision model uses the driving environment information and the strategy as input to make a decision. The decision model is optimized based on whether the decision matches the collected driving operation sequence. The cognitive model generates a strategy based on the driving results.

2. A controllable and explainable autonomous driving method according to claim 1, characterized in that: The step 1 specifically comprises the following steps: Step 1.1: Establish the cognitive model using an artificial neural network and an embedding space containing a strategy representation vector, wherein the input of the cognitive model is driving environment information, and the strategy representation vector is selected to form an output; the strategy representation vector is a vector with a length of d; Step 1.2: Using an artificial neural network to establish the decision model, the input of the decision model is the driving environment information and its corresponding strategy representation vector, and the output is the driving operation to be performed; Step 1.3: Use an artificial neural network to establish the value model, the input of the value model is the driving environment information, and the output is an evaluation value of whether the driving operation to be performed in the current decision is suitable for the current driving environment.

3. The controllable and explainable autonomous driving method according to claim 2, characterized in that: The size of the embedding space is n*d, where n represents the number of strategy representation vectors contained therein, which can be expanded with learning, and d is the length of the strategy representation vector. All strategy representation vectors are initially randomized; the driving environment information is input into the artificial neural network of the cognitive model, and a vector with a length of d is output; the similarity between the vector with a length of d and the strategy representation vector in the embedding space is calculated, and a suitable strategy vector in the embedding space is selected for output, or a combination of multiple strategy representation vectors in the embedding space is selected for output.

4. The controllable and explainable autonomous driving method according to claim 1, characterized in that: The step 2 specifically includes the following steps: Step 2.1: Use artificial neural network to establish feature extraction sub-models of cognitive model and decision model; Step 2.2: Input the driving environment information into the feature extraction sub-model, and the feature extraction sub-model outputs the corresponding environment representation vector; Step 2.3: Input the corresponding environment representation vector into the cognitive model, and the cognitive model outputs the strategy representation vector; Step 2.4: The decision model receives the corresponding environment representation vector and strategy representation vector, determines the optimal driving operation as output, and executes it.

5. The controllable and explainable autonomous driving method according to claim 1, characterized in that: The step 3 specifically comprises the following steps: Step 3.1: During runtime, the decision is controlled. At this time, the cognitive model does not accept driving environment information as input, and directly selects a specific strategy representation vector in the embedding space as output. The specific strategy representation vector is selected according to the corresponding control effect after the strategy interpretation; Step 3.2: The decision model receives the corresponding environment representation vector and strategy representation vector, determines the optimal driving operation as output, and executes it.

6. The controllable and explainable autonomous driving method according to claim 1, characterized in that: The step 4 specifically comprises the following steps: Step 4.1: Use artificial neural network to establish the feature extraction sub-model of the value model; Step 4.2: Input the driving environment information into the feature extraction sub-model, and the feature extraction sub-model outputs the corresponding environment representation vector; Step 4.3: Input the corresponding environment representation vector into the cognitive model, and the cognitive model outputs the strategy representation vector; the decision model receives the corresponding environment representation vector and the strategy representation vector, determines the optimal driving operation as output, and executes it; Step 4.4: The value model receives the corresponding environment representation vector and outputs the gap between the environment and the expectation after the driving operation is performed.

7. The controllable and explainable autonomous driving method according to claim 6, characterized in that: The step 5 specifically comprises the following steps: Step 5.1: The decision model is optimized using gradient backpropagation according to the evaluation results of the value model. The process is as follows: Step 5.1.1: Combining the optimal driving operation and the gap into a gradient; Step 5.1.2: Pass the gradient back to each component of the decision model for optimization; Step 5.2: The value model is optimized using an error back propagation algorithm based on the gap between the environment and the expectation after the driving operation is performed. The process is as follows: Step 5.2.1: The gap between the environment and the expectation after the driving operation is performed is used as the loss function; Step 5.2.2: Optimize the error back propagation algorithm using the loss function; Step 5.3: The cognitive model generates a strategy based on historical driving results. The process is as follows: Step 5.3.1: The cognitive model receives and records the driving results; Step 5.3.2: The optimization objectives of the cognitive model include two: the first is that the output vector of the artificial neural network of the cognitive model and the selected strategy representation vector have as large a similarity as possible; the second is to use the artificial neural network to construct a reconstruction sub-model, the reconstruction sub-model takes the driving operation executed by the historical decision and the driving result after execution, the driving environment information during execution, and the selected strategy representation vector as input, and the output is the optimal driving operation, and the matching of the output driving operation and the executed driving operation is used as the loss function; Step 5.3.3: Use the error back-propagation algorithm to update the artificial neural network and embedding space of the cognitive model.

8. The controllable and explainable autonomous driving method according to claim 1, characterized in that: In step 6, the cognitive model uses the driving environment information during driving to perform strategy selection as in step 2, and the decision model uses the driving environment information and the strategy to perform decision making as in step 2; the decision model uses whether the decision and the collected driving operation sequence match as an error function to perform optimization as in step 5; The cognitive model performs strategy induction generation as in step 5.

9. A controllable and explainable autonomous driving method according to any one of claims 1 to 8, characterized in that: The decision is explained by following the steps below to demonstrate interpretability: Step 7.1: Collect historical driving decision-making processes and determine the physical meaning or behavior pattern of the strategy represented by each decision representation vector in the embedding space; Step 7.2: Explain the strategy chosen during driving.

10. A controllable and explainable autonomous driving system, comprising a processor, wherein the processor is used to execute a controllable and explainable autonomous driving method as described in any one of claims 1 to 9.

Images