Unmanned aerial vehicle autonomous navigation system based on rasterized world model

Abstract

The invention discloses an unmanned aerial vehicle autonomous navigation system based on a rasterized world model, and relates to the technical field of unmanned aerial vehicle control. According to the system, observation information of an unmanned aerial vehicle is input into a trained rasterized world model through an observation acquisition module; through cooperative work of a sequence model, a multi-modal self-encoder, a hidden space dynamics predictor, a multi-modal information prediction head and a grid predictor in a rasterized world model, a cyclic variable containing historical information, a prediction hidden state vector and a prediction local grid map are provided for an agent model. And action decision making is carried out based on the data through an intelligent agent model so as to accurately output the actions of the unmanned aerial vehicle. According to the system, by introducing the rasterized world model, multi-modal observation information can be effectively fused, and a three-dimensional structure of a local environment is predicted in real time, so that the spatial perception capability of an intelligent agent model to a complex environment is enhanced, and the decision-making performance of the intelligent agent model is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of unmanned aerial vehicles, and particularly relates to an unmanned aerial vehicle autonomous navigation system based on a rasterized world model. BACKGROUND

[0002] With the rapid development of key technologies such as artificial intelligence, edge computing, communication networks, and multi-modal perception, the intelligent level of unmanned aerial vehicle systems has been significantly improved. Among them, multi-rotor unmanned aerial vehicles have been widely used in traditional scenarios such as low-altitude inspection, aerial photography, and agricultural plant protection, and have gradually expanded to fields such as air traffic, emergency rescue, and autonomous cruising in complex environments. However, the flight environment faced by unmanned aerial vehicles in the above emerging application scenarios is becoming increasingly complex, which poses technical challenges to the real-time perception accuracy, intelligent decision-making efficiency, and robust control stability of unmanned aerial vehicles. Therefore, constructing an efficient decision-making mechanism for complex dynamic environments and enabling unmanned aerial vehicles to autonomously complete navigation and collaborative tasks has become a key direction for the development of current unmanned aerial vehicle technology.

[0003] Currently, the technical route for realizing unmanned aerial vehicle autonomous navigation in dense or complex scenarios mainly includes the following two types: the first type is a navigation method based on a traditional modular structure: the autonomous navigation task of the unmanned aerial vehicle is usually divided into three relatively independent functional modules of mapping, path planning, and control, and each module executes the task in sequence according to the process. The advantage of this technical route is that the engineering interpretability is good, which is convenient for modular development, debugging, and maintenance. However, the modular method has the following disadvantages: less information interaction between modules, sensitive to unmodeled effects, and time delay in module switching. The second type is a learning-based navigation method: to overcome the limitations of the traditional modular method, this method replaces the perception, decision-making, or control modules in the traditional process with a neural network, mainly uses imitation learning or reinforcement learning algorithms, and trains the control strategy in a simulation environment using expert data or interaction data, and then migrates and deploys the trained strategy to a real unmanned aerial vehicle platform. However, the existing learning-based navigation method has the following shortcomings: poor multi-modal information fusion and utilization, poor virtual-real strategy migration capability, and low sample efficiency.

[0004] Therefore, the prior art still needs to be improved and developed. SUMMARY

[0005] The technical problem to be solved by the present application is to provide an unmanned aerial vehicle autonomous navigation system based on a rasterized world model, which aims to solve the problem that the existing learning-based navigation method cannot effectively fuse multi-modal observation information, resulting in insufficient spatial perception ability of the agent model for complex environments, and further affecting the decision-making performance.

[0006] The technical solution adopted by the present application to solve the problem is as follows:

[0007] In a first aspect, an embodiment of the present application provides an unmanned aerial vehicle autonomous navigation system based on a rasterized world model, the system comprising:

[0008] an observation acquisition module configured to acquire observation information of the unmanned aerial vehicle and input the observation information into the rasterized world model, wherein the observation information is multi-modal information;

[0009] the rasterized world model comprises a sequence model, a multi-modal autoencoder, a latent space dynamics predictor, a multi-modal information prediction head, and a raster predictor, wherein the rasterized world model is a trained model configured to generate a cycle variable at a current time step based on a latent state vector at a previous time step, an unmanned aerial vehicle action, and the cycle variable at the previous time step by using the sequence model; obtain the latent state vector at the current time step by using the multi-modal autoencoder to fuse and encode the observation information and the cycle variable at the current time step; generate a prior distribution of the latent state vector at the current time step based on the cycle variable at the current time step by using the latent space dynamics predictor, and obtain a predicted latent state vector at the current time step based on the prior distribution; and generate a predicted local raster map at the current time step based on the cycle variable at the current time step and the predicted latent state vector by using the multi-modal information prediction head and the raster predictor;

[0010] an agent model configured to generate the unmanned aerial vehicle action at the current time step based on the cycle variable at the current time step, the predicted latent state vector, and the predicted local raster map.

[0011] In an implementation manner, the system further comprises a training module configured to:

[0012] acquire observation information at a training stage by using a simulation system, and input the observation information at the training stage into the rasterized world model;

[0013] obtain a latent state vector at a previous training step by using the multi-modal autoencoder to fuse and encode the training observation information and the cycle variable at the previous training step;

[0014] generate the cycle variable at a current training step based on the latent state vector at the previous training step, the unmanned aerial vehicle action, and the cycle variable at the previous training step by using the sequence model;

[0015] generate a prior distribution of the latent state vector at the current training step based on the cycle variable at the current training step by using the latent space dynamics predictor, and obtain a predicted latent state vector at the current training step based on the prior distribution;

[0016] The multi-modal information prediction head and the grid predictor generate a predicted reward, a predicted termination signal, and a predicted local grid map of the current training step according to a loop variable of the current training step and a predicted hidden state vector;

[0017] The agent model generates a UAV action of the current training step according to the loop variable of the current training step, the predicted hidden state vector, and the predicted local grid map.

[0018] The UAV action of the current training step is input into the simulation system to realize interactive training of the agent model and the simulation system.

[0019] In an implementation manner, the observation information each includes a depth image and UAV state data; and the multi-modal autoencoder includes:

[0020] An encoder is configured to extract visual features based on the depth image through a convolutional neural network.

[0021] A multi-layer perception is configured to extract state features based on the UAV state data.

[0022] After the visual features and the state features are fused, the fused features and historical features extracted based on a loop variable are jointly subjected to discrete random encoding to obtain a hidden state vector.

[0023] In an implementation manner, the multi-modal autoencoder further includes:

[0024] An observation decoder is configured to receive the hidden state vector and the loop variable output by the encoder, and reconstruct the depth image and the UAV state data.

[0025] In an implementation manner, the grid world model generates a plurality of trajectory data in a self-recurrent manner after being trained in an end-to-end manner in a self-supervised manner.

[0026] The agent model is trained using an actor-critic algorithm according to all the trajectory data.

[0027] In an implementation manner, a loss function used for self-supervised training of the grid world model includes an observation reconstruction loss, a reward prediction loss, a termination signal prediction loss, a grid map prediction loss, a dynamic loss, and a representation loss.

[0028] In an implementation manner, a loss function used for training of the agent model using the actor-critic algorithm includes a policy loss used for optimizing an action selection policy, and a value loss used for optimizing value estimation accuracy.

[0029] In an implementation manner, the UAV action at the current time step is generated according to the cycle variable at the current time step, the predicted hidden state vector and the predicted local grid map, including:

[0030] The acceleration at the current time step is generated according to the cycle variable at the current time step, the predicted hidden state vector and the predicted local grid map through the agent model.

[0031] The UAV body z-axis direction is determined according to the acceleration at the current time step, and the yaw angle is determined according to the connecting line from the current position to the target position.

[0032] The UAV body z-axis direction is transformed into a yaw angle coordinate system with the yaw angle as the rotation angle to obtain an updated UAV body z-axis direction.

[0033] The z-axis of the world coordinate system is rotated to the updated UAV body z-axis direction to obtain a rotation axis and a rotation angle of pitch- roll rotation.

[0034] The quaternion of the rotation angle is calculated according to the rotation axis and the rotation angle, and the quaternion of rotation around the z-axis is calculated according to the rotation axis and the yaw angle.

[0035] The UAV attitude is calculated according to the quaternion of the rotation angle and the quaternion of rotation around the z-axis to obtain the UAV action at the current time step.

[0036] In a second aspect, the embodiments of the present application also provide a UAV autonomous navigation method based on a grid world model, which is applied to the UAV autonomous navigation system based on the grid world model.

[0037] The observation information of the UAV is obtained, and the observation information is input into the grid world model; the observation information is multi-modal information; the grid world model is a trained model, including a sequence model, a multi-modal autoencoder, a hidden space dynamics predictor, a multi-modal information prediction head and a grid predictor.

[0038] The cycle variable at the current time step is generated according to the hidden state vector at the previous time step, the UAV action and the cycle variable at the previous time step through the sequence model.

[0039] The hidden state vector at the current time step is obtained by fusing and encoding the observation information and the cycle variable at the current time step through the multi-modal autoencoder.

[0040] The prior distribution of the hidden state vector at the current time step is generated according to the cycle variable at the current time step through the hidden space dynamics predictor, and the predicted hidden state vector at the current time step is obtained by sampling based on the prior distribution.

[0041] The multi-modal information prediction head and the grid predictor generate a predicted local grid map of the current time step according to the loop variable and the predicted hidden state vector of the current time step;

[0042] The agent model generates a UAV action of the current time step according to the loop variable, the predicted hidden state vector and the predicted local grid map of the current time step.

[0043] In a third aspect, the embodiments of the present application further provide a terminal, which comprises a memory and one or more processors; the memory stores one or more programs; the programs contain instructions for executing the method for autonomous navigation of a UAV based on a grid world model as described above; and the processors are used to execute the programs.

[0044] The system of the present application inputs the observation information of the UAV into the trained grid world model through the observation acquisition module; the collaborative work of the sequence model, the multi-modal autoencoder, the hidden space dynamics predictor, the multi-modal information prediction head and the grid predictor in the grid world model provides the loop variable containing historical information, the predicted hidden state vector and the predicted local grid map for the agent model. The agent model makes action decisions based on these data to accurately output the UAV action. The system of the present application can effectively fuse multi-modal observation information and predict the three-dimensional structure of the local environment in real time by introducing the grid world model, thereby enhancing the spatial perception ability of the agent model for complex environments and improving its decision-making performance. BRIEF DESCRIPTION OF DRAWINGS

[0045] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed in the embodiments or the prior art description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments described in the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0046] Figure 1 is a module schematic diagram of the autonomous navigation system of the UAV based on the grid world model provided by the embodiments of the present application.

[0047] Figure 2 is a grid world model algorithm schematic diagram provided by the embodiments of the present application.

[0048] Figure 3 is a real-time simulation screen provided by the embodiments of the present application.

[0049] Figure 4 is a strategy running trajectory diagram deployed in the gazebo simulation scene provided by the embodiments of the present application.

[0050] Figure 5 This is a schematic diagram of the strategy operation provided by an embodiment of the present invention in an indoor dense obstacle scenario.

[0051] Figure 6 This is a schematic diagram of a module of an autonomous navigation system for unmanned aerial vehicles (UAVs) based on a gridded world model, provided in an embodiment of the present invention.

[0052] Figure 7 This is a schematic diagram of the terminal provided in the embodiment of the present invention. Detailed Implementation

[0053] This invention discloses an autonomous navigation system for unmanned aerial vehicles (UAVs) based on a gridded world model. To make the objectives, technical solutions, and effects of this invention clearer and more explicit, the invention is further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only for explaining the invention and are not intended to limit the invention.

[0054] Those skilled in the art will understand that, unless specifically stated otherwise, the singular forms “a,” “an,” “the,” and “the” used herein may also include the plural forms. It should be further understood that the term “comprising” as used in this specification means the presence of the stated features, integers, steps, operations, elements, and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof. It should be understood that when we say an element is “connected” or “coupled” to another element, it can be directly connected or coupled to the other element, or there may be intermediate elements. Furthermore, “connected” or “coupled” as used herein can include wireless connections or wireless coupling. The term “and / or” as used herein includes all or any units and all combinations of one or more associated listed items.

[0055] It will be understood by those skilled in the art that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. It should also be understood that terms such as those defined in general dictionaries should be understood to have the same meaning as in the context of the prior art, and should not be interpreted in an idealized or overly formal sense unless specifically defined as herein.

[0056] To address the aforementioned shortcomings of existing technologies, this invention provides an autonomous navigation system for unmanned aerial vehicles (UAVs) based on a gridded world model, such as... Figure 1 As shown, the system specifically includes:

[0057] The observation acquisition module 01 is used to acquire observation information from the UAV and input the observation information into the rasterized world model; the observation information is multimodal information.

[0058] The trained rasterized world model 02 includes: a sequence model, a multimodal autoencoder, a latent space dynamics predictor, a multimodal information prediction head, and a raster predictor. The rasterized world model is a trained model used to generate the cyclic variables for the current time step based on the latent state vector of the previous time step, UAV actions, and cyclic variables using the sequence model. The multimodal autoencoder performs fusion encoding based on the observation information and the cyclic variables of the current time step to obtain the latent state vector for the current time step. The latent space dynamics predictor generates a prior distribution of the latent state vector for the current time step based on the cyclic variables of the current time step, and samples the predicted latent state vector based on the prior distribution. The multimodal information prediction head and raster predictor generate a predicted local raster map for the current time step based on the cyclic variables and the predicted latent state vector.

[0059] Specifically, this embodiment relates to a multimodal information fusion navigation scenario. The observation information consists of raw data acquired by the UAV through various onboard sensors, relating to its own state and the surrounding environment; therefore, the UAV's observation information is multimodal. The trained rasterized world model is a mature model with environmental understanding capabilities. The UAV's observation information is input into this trained rasterized world model as the foundational data for environmental modeling, state prediction, and raster map generation. In practical applications, the UAV's observation information can be fused with data from multiple sensors, such as depth cameras and inertial navigators, to predict the 3D structure of the local environment in real time. Based on the fusion results, rapid dynamic planning and decision-making are achieved, enhancing the model's spatial perception capabilities in complex environments and improving its decision-making performance.

[0060] A rasterized world model refers to a digital model that transforms multi-source observation information from UAVs into a structured rasterized world representation and predicts the future state of the environment. Its purpose is to provide interpretable and usable environmental information for subsequent agent model decision-making. The rasterized world model mainly includes several key components: a sequence model, used to capture the temporal dependencies between the environment and actions, generating recurrent variables for the current time step using historical information to record temporal dynamic features; a multimodal autoencoder based on discrete random coding, also known as a multimodal feature variational autoencoder, used to fuse and encode multimodal observation information and recurrent variables, encoding high-dimensional observation information into latent state vectors in discrete space; a latent space dynamics predictor, used to predict environmental evolution, generating a prior distribution of latent state vectors from the recurrent variables of the current time step, and obtaining predicted latent state vectors through sampling to model the future dynamics of the environment; and a multimodal information prediction head and a raster predictor, which are the outputs of the rasterized world model, used to generate a decision-making raster map. The multimodal information prediction head predicts multimodal observation information at the current time step based on the current loop variable and the predicted latent state vector, thus verifying the effectiveness of the encoding and prediction. The grid predictor generates a predicted local grid map for the current time step based on the current loop variable and the predicted latent state vector, thereby obtaining a structured and interpretable representation of the environment. These key components each undertake different aspects of information processing, together forming a gridded world model.

[0061] For the sequence model: the loop variable refers to a variable containing historical information. If the current time step is the first time step, the loop variable of the previous time step can use a preset initial loop variable. The sequence model is a gated recurrent unit model, also known as a recurrent neural network. The sequence model uses the hidden state vector of the previous time step, the UAV action, and the loop variable as input data, and outputs the loop variable of the current time step containing historical information. For example, the hidden state vector z... t-1 Loop variable h t-1 and drone actions a t-1 As input to the sequence model, the sequence model will output a loop variable h containing historical information. t , where z, h, and a represent , respectively; t-1 and t represent the previous time step and the current time step, respectively.

[0062] For the aforementioned multimodal autoencoder: The multimodal autoencoder integrates multi-source observation information and recurrent variables into a low-dimensional, high-information latent state vector through fusion encoding, serving as an abstract representation of the overall environmental state. Taking observation information including depth images and UAV state data as an example, the multimodal autoencoder is used to fuse and encode the depth images, UAV state data, and recurrent variables, encoding the high-dimensional observation information into a latent state vector in discrete space.

[0063] In one implementation, the observation information includes: depth images and UAV status data; the multimodal autoencoder includes:

[0064] An encoder for extracting visual features from the depth image using a convolutional neural network;

[0065] State features are extracted from the UAV state data using a multilayer sensor.

[0066] After fusing the visual features and the state features, they are combined with the historical features extracted based on cyclic variables and then subjected to discrete random encoding to obtain the hidden state vector.

[0067] In summary, the encoder in a multimodal autoencoder extracts features from different types of observation information separately and then fuses them with a recurrent variable. After probabilistic and discretization processing, a latent state vector composed of one-hot encoded data is generated. Specifically, when the observation information includes depth images and UAV state data, the multimodal autoencoder uses a convolutional neural network to process the depth image to extract visual features and a multilayer perceptron to process the UAV state data to extract state features. The two are then fused, and the fused features, along with historical features extracted based on the recurrent variable, are subjected to discrete random encoding to ultimately obtain a highly fused latent state vector, serving as a compact representation of the current state.

[0068] For example, to provide the agent with sufficient navigation information, the observation space is designed as follows:

[0069] o t =[perc t ,state t ];

[0070] Among them, o t Perc represents the observation at the current time step. t The state represents the depth image at the current time step. t This indicates the drone's state at the current time step. t Including the drone's current speed v t Given target speed And the current quaternion q of drones t .

[0071] A multimodal autoencoder is used to cyclically encode high-dimensional environmental observation information and cyclic variables. The specific steps are as follows:

[0072] A multimodal autoencoder includes encoder q φ Among them, encoder q φ Includes: perc for processing depth images tConvolutional Neural Network Module and for handling drone state t Multilayer perceptron module Depth image perc t With drone state t Based on the above and After feature extraction, it is compared with the loop variable h. t The features are concatenated and fused along the feature dimension. The fused features are mapped to an intermediate variable of a preset dimension (e.g., 1024 dimensions) through a multilayer perceptron. This intermediate variable is reconstructed into a two-dimensional structure of a preset size (e.g., 32×32), and Softmax normalization is applied to the last dimension to obtain several sets (e.g., 32 sets) of discrete probability distributions. Subsequently, sampling is performed from each set of discrete probability distributions to obtain several one-hot encoded vectors equal to the number of sets (e.g., 32). All one-hot vectors are then concatenated to form a latent state vector z of a preset dimension (e.g., 1024 dimensions, equal to the dimension of the intermediate variable). t .

[0073] Furthermore, the multimodal autoencoder also includes an observation decoder, used to receive the hidden state vector and the cyclic variable output by the encoder, and reconstruct the depth image and UAV state data.

[0074] Specifically, the observation decoder p φ The primary function of the observation decoder is to verify the encoder's coding quality. By reconstructing the original observation information from the compressed hidden state vector and the cyclic variables containing temporal information, it ensures that the encoding process truly captures key information and provides supervision signals for model training. Functionally, the observation decoder can be classified as the subsequent multimodal information prediction head.

[0075] Regarding the latent space dynamics predictor: its function is to learn the evolution of the latent state vector over time. The recurrent variable is a temporal feature vector generated by the sequence model, containing all key historical information. Therefore, the recurrent variable at the current time step is used as input data for the latent space dynamics predictor to generate a prior distribution of the latent state vector at the current time step. This prior distribution describes the probability distribution of the possible values ​​of the latent state vector. By sampling this prior distribution, the distribution is transformed into a specific predicted latent state vector, thus obtaining a prediction of the latent state vector at the current time step. In practical applications, the latent space dynamics predictor can be constructed using a multilayer perceptron, taking the recurrent variable containing historical information as input and outputting a prior distribution of the latent state vector at the current time step. The obtained prior distribution is then used for sampling to predict the latent state vector at the current time step.

[0076] Regarding the multimodal information prediction head and grid predictor: During the inference phase, the loop variable of the current time step and the predicted hidden state vector are used as the input data for the multimodal information prediction head and grid predictor, so that the local grid map of the current time step can be predicted.

[0077] For example, for the multimodal information prediction head: it is constructed using a deconvolutional neural network and a multilayer perceptron, and uses the fused hidden state vector and recurrent variable as input data to jointly predict multimodal information. For the grid predictor: a multilayer perceptron is used to construct the grid predictor, using the fused hidden state vector and recurrent variable as input, to perform binary classification prediction on a 3D spatial grid map within a preset area centered on the UAV (e.g., a 4×4×2 cuboid area), providing efficient local spatial perception information to assist in the subsequent decision-making process. In the binary classification, 0 indicates no obstacles within the grid, and 1 indicates obstacles within the grid. Figure 2 As shown, CNN represents a convolutional neural network, and MLP represents a multilayer perceptron. Visual features are extracted from the depth image at time step t1 using CNN, and state features are extracted from the UAV state data at time step t1 using MLP. The hidden state vector z1 for time step t1 is generated by combining the visual features, state features, and the recurrent variable h1. The recurrent variable h2 for time step t2 is generated based on the recurrent variable h1, the hidden state vector z1, and the UAV action a1 at time step t1. Visual features are extracted from the depth image at time step t2 using CNN, and state features are extracted from the UAV state data at time step t2 using MLP. The hidden state vector z2 for time step t2 is generated by combining the visual features, state features, and the recurrent variable h2. The predicted local grid map g2 for time step t2 is generated based on the recurrent variable h2 and the hidden state vector z2.

[0078] like Figure 1 As shown, the system also includes:

[0079] The intelligent agent model 03 is used to generate the drone action for the current time step based on the cyclic variables of the current time step, the predicted hidden state vector, and the predicted local grid map.

[0080] Finally, the intelligent agent model, also known as an intelligent agent or decision-making intelligent agent, takes the current time-step cyclic variable, the hidden state vector, and the predicted local grid map as input data to generate the drone's actions for the current time step. In practical applications, the intelligent agent model can be constructed using a multilayer perceptron. Its output data is essentially three-dimensional mean and variance, which can be constructed into a multivariate Gaussian distribution. By sampling this multivariate Gaussian distribution, three-dimensional acceleration can be obtained, which is then used as the drone's control output.

[0081] In one implementation, the system further includes a training module, which is used for:

[0082] The observation information during the training phase is obtained through a simulation system, and the observation information during the training phase is input into the rasterized world model.

[0083] The multimodal autoencoder performs fusion encoding based on the training observation information and the cyclic variable of the previous training step to obtain the hidden state vector of the previous training step.

[0084] Using the sequence model, a loop variable for the current training step is generated based on the hidden state vector of the previous training deployment, the drone's actions, and the loop variable.

[0085] The latent space dynamics predictor generates a prior distribution of the latent state vector for the current training step based on the cyclic variable of the current training step, and obtains the predicted latent state vector for the current training step by sampling based on the prior distribution.

[0086] The multimodal information prediction head and grid predictor generate the predicted reward, predicted termination signal, and predicted local grid map for the current training step based on the cyclic variable and predicted hidden state vector of the current training step.

[0087] The intelligent agent model generates drone actions for the current training step based on the cyclic variable of the current training step, the predicted hidden state vector, and the predicted local grid map.

[0088] The drone's actions at the current training step are input into the simulation system to enable interactive training between the agent model and the simulation system.

[0089] Specifically, the training phase of the rasterized world model is similar to the data processing steps in the aforementioned inference phase, with the following differences:

[0090] 1. During the training phase, the observation information is provided by the simulation system.

[0091] 2. During the training phase, the decoder in the multimodal autoencoder is used to reconstruct the observations in order to calculate the loss of the reconstructed observations later.

[0092] 3. The multimodal information prediction head includes prediction modules such as a reward predictor and a termination signal predictor. During the training phase, the prediction modules in the multimodal information prediction head, which predict multimodal information such as depth map, UAV state, reward (or environmental reward), and termination signal (or cycle end signal), use the fused hidden state vector and recurrent variables as input data to jointly predict the multimodal information, thereby obtaining the predicted reward, the predicted termination signal, and the predicted local grid map. This provides high-quality data support for the model's self-supervised training and policy learning. These prediction modules can be constructed using deconvolutional neural networks and / or multilayer perceptrons.

[0093] 4. During the training phase, the drone's actions are input into the simulation system to enable interactive training between the intelligent agent model and the simulation system.

[0094] For example, firstly, a simulation system is built for navigation tasks in complex UAV scenarios, such as... Figure 3 As shown. To accelerate the convergence of the policy (i.e., the agent model), the UAV is simplified into a point mass dynamics model in the simulation environment. The point mass dynamics model models the UAV as a point mass with only three degrees of freedom in three-dimensional space, rather than a rigid body, further reducing the training difficulty and improving the training speed.

[0095]

[0096] Where a is the acceleration vector generated by the combined thrust of the four motors of the UAV, G is the gravitational acceleration vector, v is the velocity vector, and p is the position vector.

[0097] In the simulation system, the particle dynamics model is implemented based on the PyTorch deep learning framework, supporting efficient execution of UAV simulation tasks in thousands of parallel environments. For the construction of obstacle scenarios, this embodiment designs two typical environments: one is a traversing scenario composed of dense cylinders, simulating UAV flight missions in complex natural environments such as forests; the other is a narrow, dense, and deep indoor scenario composed of regular small obstacles, used to simulate the challenges of complex navigation tasks in indoor environments.

[0098] A rasterized world model is constructed between the simulation system and the agent model. The agent model inputs the agent's actions at each time step into the sequence model, and the simulation system inputs the observation information (including depth images and state data calculated based on the UAV's particle dynamics model) into the rasterized world model.

[0099] During the training phase, the functional modules of the rasterized world model include:

[0100] Multimodal encoder: z t ~q φ (z t|h t ,o t );

[0101] Observation Decoder:

[0102] Sequence model: h t =f φ (h t-1 ,z t-1 ,a t-1 );

[0103] Latent space dynamics predictor:

[0104] Reward Predictor:

[0105] Termination signal predictor:

[0106] Raster map predictor:

[0107] Wherein, the hidden state vector z t The state of environmental observations is received by a multimodal variational autoencoder. t With perc t As input data, and combined with the loop variable h t It is sampled from the posterior distribution of its output. This is achieved through the loop variable h. t It is possible to process the hidden state vector at the current time step sequentially. award Termination signal and local raster maps To make a prediction, the ^ symbol is used to distinguish the predicted value from the actual value. For the reconstructed observations obtained from decoding, According to h t Prior prediction of the distribution of the hidden state vector at the current time step. According to h respectively t and The predicted reward, termination signal, and local grid map for the current time step.

[0108] The method in this embodiment supports GPU-parallel UAV navigation and obstacle avoidance simulation, can randomly generate dense obstacle avoidance scenarios, and supports high-speed parallel simulation of thousands of UAVs.

[0109] In one implementation, the rasterized world model is trained in an end-to-end self-supervised manner and then generates several trajectory data in an autoregressive manner.

[0110] The agent model is trained using the actor-critic algorithm based on all the trajectory data.

[0111] Specifically, after self-supervised training, the rasterized world model generates predicted trajectory data through autoregression. Training the agent model using this trajectory data, specifically optimizing it with an actor-critic approach, effectively improves the agent model's decision-making performance.

[0112] For example, historical real trajectory data from eight consecutive time steps are sequentially input into a rasterized world model to obtain the initial loop variable h0 and the hidden state vector z0. Then, h0 and z0 are input into the aforementioned reward predictor, termination signal predictor, and raster map predictor to predict the reward signal. Termination signal The local grid map is g0. The policy state s0 = [h0, z0, g0] serves as the input to the agent model, which outputs the initial action a0. Finally, the sequence model takes h0, z0, and a0 as input and outputs the loop variable h1 for the next time step. The latent space dynamics predictor uses h1 as input to predict the hidden state vector for the current time step. Reward signal Termination signal Grid Map In the new cycle, h1 will be used again. Repeat the above steps to initialize the variables and obtain the new predictions. And the action a1 output by the agent model. Repeat this process for 15 steps to obtain the policy state s required for training the agent. 0:15 ,award Termination signal

[0113] The agent training uses the actor-critic framework, as shown in the following equation:

[0114] Critic:

[0115] Actor:a t ~π θ (a t |s t );

[0116] Where, π θ For the agent model, or parameterized policy function; V ψ (s t ) is a parameterized value function; γ = 0.975 is the attenuation coefficient; This represents the environmental reward generated by the rasterized world model at time step t+k. for The expectation of the dynamics between the agent model (policy action) and the gridded world model can be approximated using Monte Carlo sampling.

[0117] In one implementation, the loss function for self-supervised training of the rasterized world model includes: observation reconstruction loss, reward prediction loss, termination signal prediction loss, raster map prediction loss, dynamic loss, and representation loss.

[0118] Specifically, in the self-supervised training of the rasterized world model, the role of the loss function is to construct a supervision signal using the logic of the data itself. Observation reconstruction loss, reward prediction loss, termination signal prediction loss, raster map prediction loss, dynamic loss, and representation loss correspond to different dimensions of the model's learning objectives. These losses work together to allow the model to autonomously learn the key rules of the learning environment without relying on large amounts of manually labeled data. Observation reconstruction loss measures the error of the model in reconstructing the original observation data based on the hidden state vector, i.e., the difference between the reconstructed observation and the actual observation. In reinforcement learning or decision-making scenarios, the reward is the feedback from the environment to the agent's actions. Reward prediction loss measures the difference between the reward value predicted by the model and the actual reward value. Termination signals are specific feedback from the environment, such as the agent reaching the target point, colliding with an obstacle, or timeout indicating task termination. Termination signal prediction loss measures the difference between the termination signal predicted by the model and the actual termination signal. Raster map prediction loss measures the error between the model's predicted local raster map and the actual local raster map, which can be constructed from subsequent observation data. Dynamic loss measures the error in how the model models the dynamic changes in the environment. Representation loss optimizes the quality of representation.

[0119] For example, to guide the agent in learning navigation and obstacle avoidance strategies, this embodiment designs the reward as follows: r = 0.5r vel +4r oa +0.005r jerk +5r pos +r collide .

[0120] Where, r vel Speed ​​rewards are intended to encourage agent models to track target speeds. r oa The obstacle avoidance reward is used to guide the agent to generate appropriate lateral velocity when approaching obstacles to avoid collision risks; r jerk To ensure smooth rewards, agents are encouraged to output stable control commands, avoiding drastic policy changes in a short period of time; pos Position-keeping rewards encourage agents to achieve stable hovering at the target location; collideAs a collision penalty, a negative reward is applied when the drone collides with an obstacle. The final reward r is obtained by weighted summation of the above reward items.

[0121] The rasterized world model is trained in an end-to-end self-supervised manner, specifically using the Adam optimizer with gradient descent based on the following loss function:

[0122]

[0123] in, To observe the reconstruction loss, the loss function is the L2 loss; To reward the prediction loss, the loss function is the symlog two-hot loss; To terminate the signal prediction loss, the loss function is the binary cross-entropy loss; The loss function for raster map prediction is binary cross-entropy loss.

[0124] Dynamic loss for:

[0125]

[0126] Characterization loss for:

[0127]

[0128] Where KL[] represents the Kullback-Leibler divergence, and sg() represents the gradient stopping operator.

[0129] In one implementation, the loss function of the agent model trained using the actor-critic algorithm includes: a policy loss for optimizing the action selection strategy, and a value loss for optimizing the value estimation accuracy.

[0130] Specifically, in the actor-critic algorithm of the agent model, the actor can be viewed as a policy network responsible for outputting the probability distribution of action choices. Its goal is to learn which actions are more likely to yield high rewards in the current state. The critic can be viewed as a value network responsible for evaluating the quality of actions. Its goal is to make the value estimate as close as possible to the actual reward, providing a scoring standard for the actor's action choices. The loss function of the actor-critic algorithm helps the agent model learn to choose high-reward actions, thereby gradually learning the optimal decision-making strategy. Specifically, the policy loss optimizes the actor's action selection strategy, while the value loss optimizes the critic's value estimate, making it closer to the actual cumulative reward of the action.

[0131] For example, the training of the agent model adopts the actor-critic algorithm framework and interacts with the rasterized world model to generate the required training data such as hidden state vectors, loop variables, local raster maps, reward signals, and termination signals.

[0132] The specific training loss function for the agent model is as follows:

[0133]

[0134] in, Returns for θ:

[0135]

[0136] Where S restricts the magnitude of the gradient during policy updates, representing the current gradient magnitude. The difference between the top 5% and bottom 5% of the values ​​in the middle range. `twohot()` indicates two-hot encoding, where V... ψ (s t Using a similar structure to the reward predictor, the output is a 255-dimensional Logits. The encoding also uses two-hot encoding. In summary, the agent model can be trained using the Adam optimizer with gradient descent based on the data generated by the aforementioned rasterized world model and the designed loss function.

[0137] In one implementation, the UAV action for the current time step is generated based on the cyclic variable of the current time step, the predicted hidden state vector, and the predicted local grid map, including:

[0138] Using the intelligent agent model, the acceleration for the current time step is generated based on the cyclic variables of the current time step, the predicted hidden state vector, and the predicted local grid map.

[0139] The z-axis direction of the UAV body is determined based on the acceleration at the current time step, and the yaw angle is determined based on the line connecting the current position to the target position.

[0140] Transform the z-axis direction of the UAV body to a yaw angle coordinate system with the yaw angle as the rotation angle to obtain the updated z-axis direction of the UAV body;

[0141] Rotate the z-axis of the world coordinate system to the updated z-axis direction of the UAV body to obtain the rotation axis and rotation angle of the pitch-roll rotation;

[0142] Calculate the quaternion of the rotation angle based on the rotation axis and the rotation angle; calculate the quaternion of the rotation about the z-axis based on the rotation axis and the yaw angle.

[0143] The drone attitude is calculated based on the quaternion of the rotation angle and the quaternion of the rotation around the z-axis to obtain the drone action at the current time step.

[0144] For example, to improve policy convergence efficiency, the action space of this invention is designed as follows:

[0145] a = [a x ,a y ,a z ];

[0146] Where 'a' represents the expected acceleration of the drone in the world frame as output by the strategy, a x ,a y ,a z These are their respective components within the world system.

[0147] The system interacts with the intelligent agent model through a gridded world model module, and converts the acceleration commands output by the intelligent agent model into attitude commands required by the flight control system, thereby completing the actual deployment.

[0148] During the actual deployment phase, the full-state dynamics of the UAV need to be considered, therefore the acceleration vector output by the agent model cannot be directly used as control input. To improve inference performance on the onboard computer, the rasterized world model and agent model are exported in ONNX format, and the model inference is accelerated by the Python package onnxruntime. On an Intel N100 processor, the inference time is only 5-6ms.

[0149] The drone's hardware configuration is as follows: its core computing unit is a Risa Radxa X4 single-board computer equipped with an N100 processor, the flight control system uses an NxtPX4v2 controller based on open-source PX4 firmware, and the power part is driven by a 30A ESC module to drive a T-Motor F60 Pro V5 motor, and equipped with a five-inch T5147 propeller to provide stable thrust and maneuverability.

[0150] The overall control flow is as follows: The host computer (Radxa X4) acts as the main control platform, using the aforementioned ONNXRuntime for model inference to output control commands, and interacts with the flight controller via communication protocols such as MAVLink. The flight controller (NxtPX4v2) receives attitude control commands from the host computer, and combines them with its own inertial measurement unit (IMU), dynamic compensation system, and other sensor data to complete the underlying flight attitude calculation and control output. Subsequently, the control signal is transmitted to the brushless motor through the ESC module, driving the propeller blades to rotate, thereby realizing the thrust output and attitude adjustment of the UAV. The entire system forms a closed-loop control architecture from high-level mission planning, mid-level control allocation, and low-level motor execution. This embodiment adopts the attitude control mechanism of the PX4 flight control system, which requires converting the acceleration commands output by the agent model into corresponding thrust and attitude control quantities. Specifically, the acceleration output by the agent model and the preset yaw angle are used together to calculate the desired attitude of the UAV; among them, the modulus of acceleration is used to determine the required thrust magnitude, thereby realizing the effective conversion from agent output to actual control signals.

[0151] The conversion process is as follows: The acceleration direction is used to determine the z-axis direction of the aircraft, thereby determining the pitch and roll angles; since the z-axis direction is known, only the yaw angle needs to be determined to completely determine the UAV attitude. The direction of the line connecting the current position and the target position on the horizontal plane is used as the yaw angle reference. At the same time, combined with obstacle information on the path, the selection of the yaw angle is optimized to improve obstacle avoidance performance.

[0152] The process for determining attitude is as follows: Determine the z-axis direction of the UAV body based on the acceleration output by the agent model. Determine the yaw angle ψ = atan2(Δy, Δx) based on the line connecting the current position and the target position, where Δy and Δx are the distances from the current position to the target position relative to the x-axis and y-axis, respectively; u up Transform to the ψ coordinate system with rotation angle ψ. new =R z (-ψ)·u up , where R z (-ψ) represents the rotation matrix around the z-axis by -ψ; it rotates the world system z-axis [0,0,1] to u. new To achieve pitch-roll rotation, the rotation axis v = normalize(z × u) new Rotation angle For u new The magnitude of the sum vector along the x-axis and y-axis, For u new The projection of θ onto the z-axis is The quaternion corresponding to θ is The quaternion of rotation about the z-axis is The final drone attitude calculation is as follows:

[0153]

[0154] like Figure 3 The image shown is a real-time simulation of the system using this invention. Figure 4 and Figure 5 As shown, the model deployment effect of the present invention is that the drone can achieve a high speed of 4m / s in a narrow and dense indoor obstacle scenario.

[0155] In summary, the advantages of this invention are:

[0156] 1. Construct a UAV dynamics model and a parallel dense obstacle avoidance navigation scenario based on the PyTorch deep learning framework.

[0157] 2. A multimodal information fusion of depth image and UAV state is achieved by using convolutional neural network and multilayer perceptron. A generative world model is constructed based on variational lower bound, and local grid map prediction is achieved by using cyclic variables and hidden state vectors.

[0158] 3. A rasterized world model is trained using data obtained from interaction with simulations via a self-supervised learning paradigm. This rasterized information is then fed into the agent model to support efficient decision-making. Finally, the PX4 flight control system adjusts the strategy output, converting acceleration into thrust attitude control variables, thereby achieving virtual-real migration.

[0159] 4. A gridded model reinforcement learning method for autonomous navigation of UAVs was realized, which effectively integrates multimodal information provided by depth cameras and inertial navigation devices. By performing real-time prediction of the three-dimensional structure of the local environment, the spatial perception capability of the intelligent agent model in complex environments is enhanced, and the decision-making performance of the intelligent agent is effectively improved.

[0160] Based on the above embodiments, the present invention also provides an autonomous navigation method for unmanned aerial vehicles (UAVs) based on a gridded world model, wherein the method is applied to any of the above-described autonomous navigation systems for UAVs based on a gridded world model. Figure 6 As shown, the method includes:

[0161] Step S100: Obtain observation information from the UAV and input the observation information into the rasterized world model; the observation information is multimodal information; the rasterized world model is a trained model, including: a sequence model, a multimodal autoencoder, a latent space dynamics predictor, a multimodal information prediction head, and a raster predictor;

[0162] Step S200: Using the sequence model, generate the loop variable for the current time step based on the hidden state vector of the previous time step, the UAV action, and the loop variable;

[0163] Step S300: Using the multimodal autoencoder, the hidden state vector of the current time step is obtained by fusion encoding based on the observation information and the cyclic variable of the current time step.

[0164] Step S400: Using the latent space dynamics predictor, generate the prior distribution of the latent state vector of the current time step based on the cyclic variables of the current time step, and obtain the predicted latent state vector of the current time step based on the prior distribution.

[0165] Step S500: Using the multimodal information prediction head and the grid predictor, generate a predicted local grid map for the current time step based on the cyclic variables and predicted latent state vectors of the current time step.

[0166] Step S600: Using the agent model, generate the drone action for the current time step based on the cyclic variables of the current time step, the predicted hidden state vector, and the predicted local grid map.

[0167] Based on the above embodiments, the present invention also provides a terminal, the principle block diagram of which can be as follows: Figure 7 As shown, the terminal includes a processor, memory, network interface, and display screen connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides the environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The network interface is used for communication with external terminals via a network connection. When the computer program is executed by the processor, it implements an autonomous navigation system for the unmanned aerial vehicle (UAV) based on a rasterized world model. The display screen can be an LCD screen or an e-ink screen.

[0168] Those skilled in the art will understand that Figure 7 The schematic diagram shown is merely a partial structural diagram related to the present invention and does not constitute a limitation on the terminal to which the present invention is applied. A specific terminal may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0169] In one implementation, the terminal's memory stores one or more programs, and these programs are configured to be executed by one or more processors, and contain instructions for performing an autonomous navigation system for an unmanned aerial vehicle based on a rasterized world model.

[0170] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided by this invention can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and RAMbus dynamic RAM (RDRAM), etc.

[0171] In summary, this invention discloses an autonomous navigation system for unmanned aerial vehicles (UAVs) based on a gridded world model, relating to the field of UAV control technology. The system inputs UAV observation information into a trained gridded world model through an observation acquisition module. Through the collaborative work of modules within the gridded world model—including a sequence model, a multimodal autoencoder, a latent space dynamics predictor, a multimodal information prediction head, and a grid predictor—it provides the agent model with cyclic variables containing historical information, predicted latent state vectors, and predicted local grid maps. The agent model then makes action decisions based on this data to accurately output UAV actions. By introducing a gridded world model, this invention effectively integrates multimodal observation information, predicts the 3D structure of the local environment in real time, thereby enhancing the agent model's spatial perception capability in complex environments and improving its decision-making performance.

[0172] It should be understood that the application of the present invention is not limited to the examples above. Those skilled in the art can make improvements or modifications based on the above description, and all such improvements and modifications should fall within the protection scope of the appended claims.

Claims

1. An autonomous navigation system for unmanned aerial vehicles (UAVs) based on a gridded world model, characterized in that, The system includes: An observation acquisition module is used to acquire observation information from the UAV and input the observation information into a rasterized world model; the observation information is multimodal information. The rasterized world model includes: a sequence model, a multimodal autoencoder, a latent space dynamics predictor, a multimodal information prediction head, and a raster predictor. The rasterized world model is a trained model used to generate the cyclic variables for the current time step based on the latent state vector of the previous time step, UAV actions, and cyclic variables using the sequence model. The multimodal autoencoder performs fusion encoding based on the observation information and the cyclic variables of the current time step to obtain the latent state vector for the current time step. The latent space dynamics predictor generates a prior distribution of the latent state vector for the current time step based on the cyclic variables of the current time step, and samples the predicted latent state vector for the current time step based on the prior distribution. The multimodal information prediction head and raster predictor generate a predicted local raster map for the current time step based on the cyclic variables and the predicted latent state vector. The intelligent agent model is used to generate the drone action for the current time step based on the cyclic variables of the current time step, the predicted hidden state vector, and the predicted local grid map.

2. The UAV autonomous navigation system based on a gridded world model according to claim 1, characterized in that, The system further includes a training module, which is used for: The observation information during the training phase is obtained through a simulation system, and the observation information during the training phase is input into the rasterized world model. The multimodal autoencoder performs fusion encoding based on the training observation information and the cyclic variables of the previous training step to obtain the hidden state vector of the previous training step. Using the sequence model, a loop variable for the current training step is generated based on the hidden state vector of the previous training deployment, the drone's actions, and the loop variable. The latent space dynamics predictor generates a prior distribution of the latent state vector for the current training step based on the cyclic variable of the current training step, and obtains the predicted latent state vector for the current training step by sampling based on the prior distribution. The multimodal information prediction head and grid predictor generate the predicted reward, predicted termination signal, and predicted local grid map for the current training step based on the cyclic variable and predicted hidden state vector of the current training step. The intelligent agent model generates drone actions for the current training step based on the cyclic variable of the current training step, the predicted hidden state vector, and the predicted local grid map. The drone's actions at the current training step are input into the simulation system to enable interactive training between the agent model and the simulation system.

3. The UAV autonomous navigation system based on a gridded world model according to claim 1 or 2, characterized in that, The observation information includes: depth images and UAV status data; the multimodal autoencoder includes: An encoder for extracting visual features from the depth image using a convolutional neural network; State features are extracted from the UAV state data using a multilayer sensor. After fusing the visual features and the state features, they are combined with the historical features extracted based on cyclic variables and then subjected to discrete random encoding to obtain the hidden state vector.

4. The UAV autonomous navigation system based on a gridded world model according to claim 3, characterized in that, The multimodal autoencoder also includes: An observation decoder is used to receive the hidden state vector and cyclic variables output by the encoder and reconstruct the depth image and UAV state data.

5. The UAV autonomous navigation system based on a gridded world model according to claim 2, characterized in that, The rasterized world model is trained in an end-to-end self-supervised manner and then generates several trajectory data in an autoregressive manner. The agent model is trained using the actor-critic algorithm based on all the trajectory data.

6. The UAV autonomous navigation system based on a gridded world model according to claim 5, characterized in that, The loss functions for self-supervised training of the rasterized world model include: observation reconstruction loss, reward prediction loss, termination signal prediction loss, raster map prediction loss, dynamic loss, and representation loss.

7. The UAV autonomous navigation system based on a gridded world model according to claim 5, characterized in that, The loss function of the agent model trained using the actor-critic algorithm includes: a policy loss for optimizing the action selection strategy, and a value loss for optimizing the value estimation accuracy.

8. The UAV autonomous navigation system based on a gridded world model according to claim 1, characterized in that, Based on the current time step's loop variables, predicted hidden state vectors, and predicted local grid map, generate the UAV actions for the current time step, including: Using the intelligent agent model, the acceleration for the current time step is generated based on the cyclic variables of the current time step, the predicted hidden state vector, and the predicted local grid map. The z-axis direction of the UAV body is determined based on the acceleration at the current time step, and the yaw angle is determined based on the line connecting the current position to the target position. Transform the z-axis direction of the UAV body to a yaw angle coordinate system with the yaw angle as the rotation angle to obtain the updated z-axis direction of the UAV body; Rotate the z-axis of the world coordinate system to the updated z-axis direction of the UAV body to obtain the rotation axis and rotation angle of the pitch-roll rotation; Calculate the quaternion of the rotation angle based on the rotation axis and the rotation angle; calculate the quaternion of the rotation about the z-axis based on the rotation axis and the yaw angle. The drone attitude is calculated based on the quaternion of the rotation angle and the quaternion of the rotation around the z-axis to obtain the drone action at the current time step.

9. An autonomous navigation method for unmanned aerial vehicles (UAVs) based on a gridded world model, characterized in that, The method is applied to the UAV autonomous navigation system based on a rasterized world model as described in any one of claims 1-8, and the method includes: The observation information of the UAV is acquired and input into the rasterized world model; the observation information is multimodal information; the rasterized world model is a trained model, including: a sequence model, a multimodal autoencoder, a latent space dynamics predictor, a multimodal information prediction head and a raster predictor; Using the sequence model, the loop variable for the current time step is generated based on the hidden state vector of the previous time step, the UAV action, and the loop variable. The multimodal autoencoder performs fusion encoding based on the observation information and the cyclic variables of the current time step to obtain the hidden state vector of the current time step. The latent space dynamics predictor generates a prior distribution of the latent state vector of the current time step based on the cyclic variables of the current time step, and obtains the predicted latent state vector of the current time step based on the prior distribution. Using the multimodal information prediction head and the grid predictor, a predicted local grid map for the current time step is generated based on the cyclic variables and the predicted latent state vector at the current time step. Using an intelligent agent model, the drone actions for the current time step are generated based on the cyclic variables of the current time step, the predicted hidden state vector, and the predicted local grid map.

10. A terminal, characterized in that, The terminal includes a memory and one or more processors; the memory stores one or more programs; the programs contain instructions for executing the UAV autonomous navigation method based on a gridded world model as described in claim 9; the processors are used to execute the programs.

Images