A model predictive control system and method with heterogeneous neural network inlay collaboration

Abstract

The application discloses a kind of isomorphic neural network embedded collaborative model predictive control system and method, belong to the cross field of automatic control and artificial intelligence.System includes Koopman neural network prediction model, realizes high-precision, physical consistent system identification;Differentiable planning network rolling optimizer, realize microsecond level optimal control solution with constraint;Hybrid drive adaptive neural network observer, realize state estimation and mutation fault tolerance;Intelligent reference trajectory generator, realize continuous smooth, task adaptive trajectory planning.Method includes: trajectory generator generates reference trajectory, observer outputs modified state estimation, prediction model predicts future output, optimizer solves optimal control quantity and acts on controlled object.The application solves the pain points of traditional MPC modeling difficulty, slow solution, weak robustness, trajectory rigid, significantly improves the precision, real-time, robustness and adaptability of control system, and can be widely applied to automatic driving, robot and other high dynamic strong constraint scene.

Description

Technical Field

[0001] This invention belongs to the field of interdisciplinary technology of automatic control and artificial intelligence, and specifically relates to a model predictive control system and method with heterogeneous neural network embedded in collaboration. Background Technology

[0002] Model predictive control (MPC) has become a core algorithm in fields such as autonomous driving due to its explicit handling of constraints and rolling optimization. In autonomous driving, MPC achieves following or cruising at the desired speed by adjusting throttle opening and braking pressure, lane keeping, lane changing, and cornering by controlling the front wheel steering angle, plans overtaking trajectories and executes emergency obstacle avoidance maneuvers, and coordinates the cooperative driving of multiple vehicle platoons, while simultaneously meeting vehicle dynamics constraints, actuator limits, and safe distance requirements from surrounding obstacles. The traditional MPC framework consists of four parts: a predictive model, rolling optimization, feedback correction, and a reference trajectory.

[0003] Traditional MPC faces three main technical bottlenecks: (1) Predictive models are difficult to accurately describe strongly nonlinear systems, and pure data-driven models lack physical consistency constraints, resulting in poor generalization ability under small sample conditions. (2) Constrained quadratic programming solutions have large computational loads, resulting in real-time bottlenecks in highly dynamic scenarios and making them difficult to deploy on edge computing platforms. (3) Traditional state observers are not robust enough to sudden disturbances and sensor failures, and their state estimation accuracy is limited. Existing technologies attempt to introduce neural networks into MPC, but they suffer from common defects such as local substitution and isolated embedding. Chinese patent document CN121209571A discloses an agent model predictive control method based on Koopman neural network, which uses Koopman neural network to establish a predictive model for the agent. This scheme only uses Koopman neural network for the predictive model and does not involve the coordinated transformation of optimization solution and feedback observation links. Under dynamic environment control, real-time performance and stability still need to be improved.

[0004] To address the aforementioned issues, this solution proposes a systematic MPC architecture that embeds predictive modeling, rolling optimization, feedback correction, and reference trajectory generation into various neural networks, with each function supporting and cooperating with the others. This substantially improves the accuracy, real-time performance, and robustness of the control system. Summary of the Invention

[0005] This invention aims to solve the following technical problems:

[0006] 1. Break through the performance bottleneck of existing single-stage embedded neural network MPC and realize the full-process neural networkization of the four stages of prediction, optimization, correction and reference trajectory.

[0007] 2. To address the issues of pure data-driven prediction models failing to reflect physical laws and exhibiting poor generalization ability with small sample sizes.

[0008] 3. Addresses the issue of high computational latency in constrained MPC online optimization, making it difficult to deploy on edge computing platforms.

[0009] 4. Address the problem of insufficient robustness of traditional observers to sudden disturbances and sensor failures.

[0010] 5. Resolves the issue of being unable to dynamically match task intent and outputting unevenly due to using a fixed smoothing coefficient for the reference trajectory.

[0011] The present invention achieves the above objectives using the following technical solutions:

[0012] A heterogeneous neural network embedded collaborative model predictive control system includes four modules: prediction, optimization, observation, and planning. Each module is independent yet coupled with others, and together they form a differentiable closed-loop control system.

[0013] The prediction module is a Koopman neural network prediction model based on physical information fusion, consisting of an encoder, a Koopman linear dynamics layer, and a decoder. The encoder elevates the nonlinear system state to a high-dimensional linear space, the Koopman linear dynamics layer performs linear state transitions in the latent space, and the decoder maps the latent states to the original state space. The loss function of the prediction model consists of three weighted components: reconstruction loss, multi-step prediction loss, and physical information loss. The physical information loss uses the system residual mechanism model or the output of conservation as a regularization constraint, forcing the prediction trajectory of the Koopman neural network to conform to basic physical laws, thus maintaining its extrapolation ability even with sparse training data.

[0014] The Koopman neural network is implemented using a differentiable neural network, with its encoder and decoder being fully connected networks or convolutional networks, and the linear dynamics matrix K being a trainable parameter.

[0015] The optimization module is a rolling optimizer based on a differentiable programming network. It receives the future state sequence and reference trajectory from the prediction module and uses a neural network parameterized control strategy. The optimizer embeds constraints through a differentiable quadratic programming layer. It is trained end-to-end, with the loss function defined as a weighted sum of the original MPC objective function and the penalty for violating control constraints. After training convergence, the optimizer only requires one forward propagation during deployment to output the approximate optimal control quantity, eliminating the need for iterative QP problem solving.

[0016] Differentiable programming networks are implemented based on the OptNet architecture or the cvxpylayers framework, and support automatic differentiation of inequality constraints and boundary constraints.

[0017] The observation module is a hybrid-driven adaptive neural network observer, which receives the actual output of the system. Control input Output state estimate and disturbance estimates The observer employs a dual-channel heterogeneous architecture. The first channel is a nonlinear error observer based on RBFNN, used to approximate unmodeled dynamics. The second channel is a fault detector based on residual Gaussian process regression, used to detect sudden sensor faults online and reconstruct the correct sensor output. The outputs of the two channels are adaptively weighted and fused, and then input into the prediction model as the corrected state at the current time step.

[0018] The residual Gaussian process regression uses the Matern 5 / 2 covariance function, with the residual sequence of the actual output and the output of the prediction model within the sliding window as training data, and the covariance function parameters are updated in real time.

[0019] The planning module is an intelligent reference trajectory generator based on the KAN diffusion strategy. It receives task instructions and environmental information and outputs future trajectory information. A smooth reference trajectory sequence is generated step by step. Its core architecture employs a diffusion probability model, replacing the convolutional layers in traditional UNet with KAN layers. The KAN layers use learnable spline functions as activation primitives, possessing inherent continuous differentiability, effectively suppressing the staggered, abrupt changes in the output trajectory of traditional diffusion strategies. The reference trajectory generator is pre-trained using behavior cloning and fine-tuned with action consistency regularization.

[0020] The prediction module, optimization module, observation module, and planning module are connected sequentially to form a closed-loop control system. The system executes the following steps in each control cycle.

[0021] S1: The planning module generates a smoothing reference trajectory sequence based on the current task objective. And output it to the optimization module.

[0022] S2: The observation module receives the actual output. Output correction status To the prediction module.

[0023] S3: Prediction module with Given the initial state, and combining it with candidate control sequences, predict the future output. This information is then fed back to the optimization module.

[0024] S4: Differentiable programming networks aim to quickly solve for the current optimal control input by predicting the deviation between the output and the reference trajectory. And output it to the controlled object or the executing mechanism.

[0025] S5: Will Acting on the controlled object or the executing agency, ,return .

[0026] Compared with the prior art, the present invention has the following significant advantages:

[0027] 1. Significantly improved control accuracy: This invention employs a physically constrained Koopman neural network prediction model, introducing a physical information regularization term from a priori mechanism model into the loss function. This allows the model to still follow physical laws for prediction even with sparse training data, effectively improving prediction accuracy. The observer based on residual GPR can compensate for model mismatch errors online, further enhancing the accuracy of state estimation. This invention can significantly reduce lateral errors in path tracking and improve control accuracy.

[0028] 2. Significantly improved real-time performance: This invention uses a differentiable programming network to replace the traditional QP solver, transforming the online optimization problem into a feedforward neural network computation, which greatly reduces online computation time and can meet the real-time control requirements of high-dynamic scenarios such as drones and humanoid robots.

[0029] 3. Enhanced robustness and fault tolerance: This invention adopts a dual-channel heterogeneous observer architecture. When a sensor experiences a sudden failure, the residual GPR channel can quickly detect and reconstruct the correct signal, enabling the system to maintain stable control. The fault recovery time is far superior to that of traditional observers.

[0030] 4. Improved motion smoothness and task adaptability: This invention uses the KAN diffusion strategy to generate reference trajectories. By utilizing the spline basis function characteristics of the KAN layer, the output trajectory has inherent continuous differentiability, which effectively suppresses the trajectory jerking that is easy to produce in traditional methods, and improves motion smoothness and task execution success rate.

[0031] 5. Enhanced system integration and portability: The four modules described in this invention are all differentiable, enabling end-to-end optimization training from sensor input to control output, supporting Simulation to Reality migration from simulation environment to physical platform, and providing good deployment flexibility. Attached Figure Description

[0032] Figure 1 This is a system overall architecture block diagram of Embodiment 1 of the present invention;

[0033] Figure 2 This is a schematic diagram of the physical information fusion Koopman neural network prediction model in Embodiment 1 of the present invention;

[0034] Figure 3 This is a comparison diagram of the training and deployment process of the differentiable programming network optimizer in Embodiment 1 of the present invention;

[0035] Figure 4This is an internal structure diagram of the hybrid-driven adaptive neural network observer in Embodiment 1 of the present invention;

[0036] Figure 5 This is a network architecture diagram of the intelligent reference trajectory generator based on the KAN diffusion strategy in Embodiment 1 of the present invention. Detailed Implementation

[0037] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. The specific embodiments of this invention are only for explaining the invention and are not intended to limit the scope of protection of this invention.

[0038] Example 1: Path tracking control applied to autonomous vehicles.

[0039] like Figure 1 As shown, this invention is applied to path tracking control of a certain type of passenger vehicle. The overall system architecture includes: an intelligent reference trajectory generator 100, a hybrid drive adaptive neural network observer 200, a physical information fusion Koopman neural network prediction model 300, a differentiable programming network rolling optimizer 400, and a controlled object 500. The controlled object 500 adopts a two-degree-of-freedom bicycle model as the prior mechanism model.

[0040] Module 1 (Prediction Module): Koopman Neural Network Prediction Model 300 based on physical information fusion.

[0041] like Figure 2 As shown, the physical information fusion Koopman neural network prediction model 300 includes an encoder 310, a linear dynamics layer 320, a decoder 330, a multi-step prediction unfolding module 340, a physical information constraint module 350, and a total loss function calculation module 360.

[0042] Encoder 310 (denoted as A three-layer fully connected neural network is used, with the input layer dimension equal to the state dimension. The state has 5 dimensions. In this example, the state is 5-dimensional, representing the vehicle's longitudinal velocity, lateral velocity, yaw rate, lateral position, and heading angle. The hidden layer has 128 and 64 neurons respectively, and the output layer has a dimension of 64. The activation function used is ReLU. Encoder 310 will output the current state. Mapped to hidden state .

[0043] 320 parameter matrix of linear dynamics layer , Initialized as a random matrix, it serves as a trainable parameter in the optimization. This layer performs a linear state transition in the latent space as follows: .

[0044] Decoder 330 (denoted as A three-layer fully connected network symmetrical to encoder 310 is used to hide the hidden state. Map back to the original state space and output the predicted state. .

[0045] The multi-step prediction unfolding module 340 obtains the future by iterating through the above process. Predicted sequence of steps: , That is, the system predicts the output. .

[0046] like Figure 2 As shown, the physical information constraint module 350 includes a mechanism model prior 351 and a physical information loss calculation 352. The mechanism model prior 351 calculates based on the system residual mechanism model. Physical information loss calculation 352 calculation .

[0047] The total loss function calculation module 360 ​​calculates the total loss. = + in To rebuild the losses, To predict losses in multiple steps, , .

[0048] Module 2 (Optimization Module): Rolling Optimizer for Differentiable Programming Networks.

[0049] like Figure 3 As shown, the differentiable programming network rolling optimizer 400 includes a training phase 410 and a deployment phase 420, both of which share the same network architecture but have different parameter states.

[0050] Training phase 410, policy network A two-layer fully connected network is used, with 64 neurons in the hidden layer and the output dimension being the control variable u. The dimension. Differentiable QP layer 413 is implemented based on the cvxpylayers framework and supports controllable magnitude constraints on front wheel steering angle. acceleration Controlling the rate of change of rotation angle by incremental constraints rate of change of acceleration and output state constraints, lateral position deviation During the training phase, a conventional QP solver 411 was used to generate the solution under random initial state sampling. The optimal trajectory is used as supervision data to train the policy network and the differentiable QP layer using behavior cloning. The training loss function is a weighted sum of mean squared error and penalty for violation of control constraints.

[0051] Deployment phase 420: After training, freeze the parameters of the policy network 421 and the differentiable QP layer 422. State After input, only one forward propagation is needed to output the control quantity. Control cycle setting The measured single-step computation time on the NVIDIA Jetson Orin platform is as follows: Compared to traditional MPC This step increases the calculation speed by more than 10 times.

[0052] Constraints are uniformly input through constraint input module 430 and are embedded into the forward and backward propagation process of the differentiable QP layer using the Lagrange multiplier method.

[0053] Module 3 (Observation Module): Hybrid-driven adaptive neural network observer.

[0054] like Figure 4 As shown, the hybrid-driven adaptive neural network observer 200 includes an input signal receiving module 210, an RBFNN channel 220, a residual GPR channel 230, and an adaptive weighted fusion module 240.

[0055] Input signal receiving module 210 receives control input. Actual output and predicted output .in and Input to RBFNN channel 220, and Input to GPR channel 230.

[0056] The RBFNN channel 220 includes a radial basis function layer 221 and a Lyapunov weight update layer 222. The RBFNN observer uses 15 radial basis centers, uniformly distributed within the expected range of the state space, with a basis width of... The online weight update law is designed based on Lyapunov stability theory to ensure the convergence of estimation errors. The RBFNN channel output state estimates are... .

[0057] The residual GPR channel 230 includes a residual calculation module 231, a Matern covariance calculation layer 232, and a fault detection module 233. The residual calculation module 231 calculates... The Matern covariance calculation layer 232 uses the Matern 5 / 2 covariance function, with a sliding window length of... ,Every Update the hyperparameters of the Gaussian process once. The fault detection module 233 sets the fault detection threshold to... When the confidence level is below a threshold, it is determined to be a sensor fault. GPR channel output state estimate. and fault confidence.

[0058] The adaptive weighted fusion module 240 dynamically fuses two-channel outputs: RBFNN weights under normal operating conditions. GPR weight The weighted formula is: When the fault diagnosis module 233 determines that the sensor is faulty, the RBFNN channel weights are reset to zero, and the system completely switches to the GPR channel. The final state estimation output terminal 250 outputs the corrected state. Enter to Figure 1 The physical information fusion Koopman neural network prediction model 300.

[0059] Module 4 (Planning Module): Intelligent Reference Trajectory Generator Based on KAN Diffusion Strategy.

[0060] like Figure 5 As shown, the intelligent reference trajectory generator 100 based on the KAN diffusion strategy includes a forward diffusion process and a reverse denoising network. The forward diffusion process gradually moves towards the true trajectory. Add Gaussian noise, after Step to obtain pure noise The diffusion formula for each step is: ,in The number of diffusion steps in this embodiment .

[0061] The inverse denoising network employs a noise prediction network. Internally, it includes an input layer, an RWKV temporal feature extraction layer 121, a GroupKAN layer 122, and an output layer. The RWKV temporal feature extraction layer 121 handles the temporal dependencies of the trajectory sequence. The GroupKAN layer 122 contains spline primitives 123 and trainable grid parameters 124, performing group-wise spline basis nonlinear mapping. The KAN layer has 5 spline grids, a spline order of 3, and the grid parameters are trainable variables. The core formula for the reverse denoising process is:

[0062] .

[0063] The training objective is to minimize the loss function. The training dataset consists of 20,000 trajectory segments from the publicly available Highway Driving Dataset. Fine-tuning is performed using action consistency regularization loss, with weight coefficients... .

[0064] During the deployment phase, the trajectory generator 100 executes the task instructions. Generate the future through a reverse denoising process Step-softening reference trajectory sequence and output to Figure 1 The differentiable programming network rolling optimizer 400 is used in this embodiment. .

[0065] Theoretical analysis of this embodiment was conducted using the CarSim / Simulink co-simulation platform. The results show that the present invention can significantly reduce the lateral error of path tracking, make the control variable change smoother, and the single-step calculation time is much shorter than the control cycle, thus meeting the requirements of real-time control.

[0066] Example 2: Application to autonomous berthing control of unmanned surface vessels.

[0067] This invention is applied to the autonomous berthing task of a certain type of underactuated unmanned surface vessel. The controlled object is a 1:70 scale model of the CyberShip II, 1.255m long, 17.6kg in mass, and equipped with a single propeller and single rudder propulsion system.

[0068] Module 1: Koopman Neural Network Prediction Model

[0069] encoder A three-layer fully connected neural network is used, with the input state consisting of six dimensions: position and velocity (three degrees of freedom), including sway, roll, and pitch. The hidden layer has 64 and 32 neurons respectively, and the output layer has a dimension of 32. The ReLU activation function is used. A linear dynamics layer is also employed. , The control inputs are propeller speed and rudder angle. Decoder Hidden state Mapping back to a 6-dimensional state. The physical information loss term is based on the MMG separable modeling approach, using the residuals of the Abkowitz nonlinear hydrodynamic model as constraints to implicitly encode ship maneuvering behavior. In the total loss function... , .

[0070] Module 2: Differentiable Programming Network Rolling Optimizer

[0071] Policy Network It adopts a two-layer fully connected network with a 64-dimensional hidden layer and outputs control variables. For propeller speed and rudder angle. Constraints include rudder angle limitation. Rudder speed limit propeller speed limit And the rudder effectiveness constraint at low speeds. The control cycle is set to... To accommodate the large inertia characteristics of ships, the optimizer is deployed on an industrial control computer, and the single-step calculation time is measured in practice. .

[0072] Module 3: Hybrid-driven adaptive neural network observer:

[0073] The RBFNN channel receives three signals: GPS position, attitude and heading, and odometer speed. It has 20 center points and a base width of [missing information]. The residual GPR channel receives rudder angle feedback and propeller speed feedback to detect actuator faults. The fusion module dynamically adjusts the weights based on speed: the RBF weight is 0.8 at high speeds and increases to 0.6 when berthing at low speeds. The fault detection threshold is 0.9, and the module automatically switches to GPR dead reckoning when GPS signal is lost.

[0074] Module 4: KAN Diffusion Strategy Reference Trajectory Generator

[0075] The reference trajectory is dynamically generated based on electronic chart information, real-time wind, wave, and current data, and berth geometric constraints. The mission instruction includes the target berth coordinates. Expected heading Maximum berthing speed Diffusion steps The spline grid has 8 grid points and a spline order of 3. The training dataset consists of historical AIS trajectory data from actual ports such as Dalian Port and Qingdao Port, as well as berthing trajectories generated by ship maneuvering simulators, totaling 30,000 trajectory segments.

[0076] Simulation verification and expected results:

[0077] Analysis was conducted using MATLAB / Simulink and a marine environment simulator platform. Environmental disturbances were set to sea state 6, with a wind speed of 12 m / s, a significant wave height of 4.5 m, a current velocity of 0.5 m / s, and a current direction at a 30° angle to the berth centerline. Results show that the maximum lateral position error during berthing can be controlled within 0.3 m, equivalent to approximately 21 m on a real ship. Compared to traditional nonlinear model predictive control, the scale error is 0.9 m, representing a reduction of approximately 66.7%. The peak rate of change of rudder angle is expected to be reduced by 41.3% compared to traditional nonlinear model predictive control. Even with a 5-second GPS signal interruption, the observer can maintain dead reckoning, with an expected position drift of ≤1.2 m. The single-step calculation time is expected to be ≤25 ms, meeting the real-time requirement of a 100 ms control cycle.

[0078] In addition to the above embodiments, the present invention can also be applied to fields such as quadcopter drone trajectory tracking, bipedal robot balance control, and smart grid load frequency control. Their common feature is that they all adopt the collaborative architecture of the four neural network embedded modules.

[0079] While the specific embodiments of the present invention have been described in detail above, the present invention is not limited to the above embodiments. Within the scope of knowledge possessed by those skilled in the art, various changes can be made without departing from the spirit of the present invention, and modifications or variations without creative effort are still within the protection scope of the present invention.

Claims

1. A model predictive control system with heterogeneous neural network embedded in collaboration, characterized in that, include: Prediction module, optimization module, observation module, and planning module: The prediction module is a physical information fusion Koopman neural network prediction model, which is used to predict the future output of the system based on the current state and control input. Its loss function includes physical information constraint terms based on prior mechanism models or conservation laws. The optimization module is a differentiable programming network rolling optimizer, which is used to solve the constrained optimal control sequence in a single forward propagation manner. The optimizer includes a parameterized control policy network and a differentiable quadratic programming layer. The observation module is a hybrid-driven adaptive neural network observer used to estimate the system state online and detect sensor faults. The observer includes a radial basis function neural network channel and a residual Gaussian process regression channel. The planning module is an intelligent reference trajectory generator based on the KAN diffusion strategy, which is used to generate smooth reference trajectories according to task instructions and environmental perception. The generator uses the Kolmogorov-Arnold network as the inverse denoising backbone network of the diffusion probability model. The Koopman neural network prediction model, the differentiable programming network rolling optimizer, the hybrid-driven adaptive neural network observer, and the intelligent reference trajectory generator based on the KAN diffusion strategy are sequentially connected to form a closed-loop control system. The prediction module, optimization module, observation module, and planning module are all differentiable architectures, supporting end-to-end joint optimization.

2. The system according to claim 1, characterized in that, The loss function of the Koopman neural network prediction model is specifically expressed as follows: ； in To rebuild the losses, , These are the weighting coefficients. To predict losses in multiple steps, The physical information loss is calculated as the sum of squared deviations between the output of the system residual mechanism model at a selected state point and the output of the Koopman model.

3. The system according to claim 1, characterized in that, The differentiable programming network optimizer adopts a joint training strategy of behavior cloning and constraint violation punishment. The training data comes from the optimal trajectory generated by the traditional numerical optimization solver under the initial state sampling of Gaussian distribution. During the deployment phase, the differentiable programming network optimizer is completely separated from the solver and only performs forward computation of the neural network.

4. The system according to claim 1, characterized in that, The hybrid-driven adaptive neural network observer includes: The first channel is based on a radial basis function neural network, and its weights adopt an online adaptive update law based on Lyapunov stability theory; The second channel is based on residual Gaussian process regression and uses the Matern 5 / 2 covariance function to update hyperparameters online using the output residual sequence within the sliding window as training data. The fusion module dynamically weights the output of the two channels based on the confidence level of the fault detection.

5. The system according to claim 1, characterized in that, The intelligent reference trajectory generator based on the KAN diffusion strategy includes: During the forward diffusion process, Gaussian noise is gradually added to the actual trajectory; In the reverse denoising process, the RWKV layer is used for temporal feature extraction, and the GroupKAN layer is used for nonlinear mapping of spline basis functions group by group. The number of spline grids in the GroupKAN layer is adjustable, and the grid parameters are trainable variables.

6. The system according to any one of claims 1 to 5, characterized in that, The prediction module, optimization module, observation module, and planning module are jointly trained end-to-end through a differentiable programming framework. The total loss function of the joint training is a weighted sum of the loss functions of each module.

7. A heterogeneous neural network embedded collaborative model predictive control method, applied to the system described in any one of claims 1 to 6, characterized in that, Includes the following steps: Step 1: The planning module receives the task instruction and generates the future... Step-softening reference trajectory sequence ; Step 2: The observation module receives the actual output at the current moment. Output the corrected state estimate ; Step 3: The prediction module uses Using the initial state and candidate control sequences provided by the optimizer, calculate the predicted output. ; Step 4: The optimization module aims to minimize the deviation between the predicted output and the reference trajectory, and solves for the optimal control quantity at the current time step using a differentiable programming layer. ; Step 5: Output to the controlled object and update the time index. Return to step 1.

8. The method according to claim 7, characterized in that, In step 4, the constraints of the differentiable programming layer include control magnitude constraints, control increment constraints, and output state constraints. The constraints are embedded into the forward and backward propagation processes of the differentiable programming layer using the Lagrange multiplier method.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method as described in claim 7 or 8.

10. An autonomous vehicle controller, characterized in that, Includes the system described in any one of claims 1 to 6.

Images