An omnidirectional flying robot high-precision hovering method and system

Abstract

The application relates to the technical field of unmanned plane control, and discloses an omnidirectional flying robot high-precision hovering system and method, which comprises the following steps: collecting the connection state of a module assembly, allowing flying action when it is determined to be normal; collecting the GNSS signal strength of a positioning sensor, determining a positioning decision mode according to the GNSS signal strength, and generating flight attitude control data according to the positioning decision mode; adjusting the flight attitude of the flying robot according to the flight attitude control data and based on the closed-loop PID control of a harmonic reducer and a servo motor; obtaining predicted wind speed data to determine a disturbance torque, and feeding forward compensating the flight attitude according to the disturbance torque; controlling the central double rotors of the flying robot to continuously run, and controlling the start and stop of the inner and outer rotors of the flying robot based on Q-learning. The application guarantees the wind resistance hovering performance, improves the energy utilization rate, and enhances the environmental adaptability, wind resistance stability and endurance of the aircraft.

Description

Technical Field

[0001] The present invention relates to the field of unmanned aerial vehicle (UAV) control technology, and in particular to a high-precision hovering method and system for an omnidirectional flying robot. Background Art

[0002] With the increasing application of unmanned aerial vehicles (UAVs) in wind turbine facility inspection, high-voltage transmission line monitoring, and power grid troubleshooting, the ability of flying robots to achieve high-precision hovering and stable attitude control in complex, high-altitude, and multi-disturbance environments has become a key technical requirement. For example, during wind turbine blade inspection, the UAV must remain stationary for extended periods within three meters of the blade edge to facilitate visual or ultrasonic inspection. Similarly, when operating within high-voltage transmission line corridors, hovering stability is directly impacted by factors such as terrain obstruction, electromagnetic interference, and strong wind disturbances. Traditional flight control systems, relying solely on GNSS (Global Navigation Satellite System) positioning, struggle to ensure the aircraft's spatial stability and hovering accuracy. Furthermore, environmental wind disturbances, power system response delays, and inadequate energy management further limit the ability of existing UAVs to maintain sustained hovering in complex environments.

[0003] In existing technology, most flying robots rely on fixed-direction thrust systems, using a traditional quadrotor structure coupled with a simple PID controller for attitude and position control. However, this type of structure suffers from issues such as unadjustable thrust direction, slow dynamic response, weak wind resistance, and accumulated positioning errors in GNSS-deactivated environments. Furthermore, during flight control, the continuous operation of all rotors wastes energy, limits flight time, and prevents flexible optimization of thrust resources based on the environment and flight load.

[0004] While some technologies have incorporated visual-inertial navigation, UWB (Ultra Wide Band) ranging systems, or dynamic thrust allocation strategies to improve hovering accuracy and endurance, practical limitations remain. For example, single-use visual or inertial navigation systems lack reliability in feature-sparse or dynamic scenarios, making them incapable of independently supporting high-precision hovering. Traditional direct motor drive suffers from response lag and transmission slack, making centimeter-level stable control difficult. Furthermore, existing energy management methods often employ fixed strategies that fail to adjust rotor start and stop in real time based on flight conditions, resulting in low energy efficiency and limited improvements in endurance.

[0005] Therefore, it is necessary to design a high-precision hovering method and system for an omnidirectional flying robot to solve the problems existing in current technology. Summary of the Invention

[0006] In view of this, the present invention proposes a high-precision hovering method and system for an omnidirectional flying robot, aiming to solve the problems of slow dynamic response speed, weak wind resistance and low energy efficiency in the current hovering of flying robots.

[0007] In one aspect, the present invention provides a high-precision hovering method for an omnidirectional flying robot, comprising:

[0008] Collecting the connection status of module components in the flying robot, and allowing flight actions when the module components are determined to be connected normally. The module components include positioning sensors, harmonic tilt-rotors, IMUs (Inertial Measurement Units), binocular vision systems, UWB, and millimeter-wave radars;

[0009] When the takeoff action is completed and it is determined to hover, the GNSS signal strength of the positioning sensor is collected, a positioning decision mode is determined based on the GNSS signal strength, and flight attitude control data is generated based on the positioning decision mode. The positioning decision mode includes a GNSS-EKF (Extended Kalman Filter) fusion mode and a SLAM+UWB particle filter fusion mode.

[0010] Adjusting the flight attitude of the flying robot according to the flight attitude control data and based on closed-loop PID control of a harmonic reducer and a servo motor;

[0011] Obtaining predicted wind speed data according to the IMU and the millimeter-wave radar, determining a disturbance torque according to the predicted wind speed data, and performing feedforward compensation on the flight attitude according to the disturbance torque;

[0012] The central dual rotors of the flying robot are controlled to operate continuously, and the start and stop of the middle and outer rotors of the flying robot are controlled based on Q-learning.

[0013] Furthermore, collecting the GNSS signal strength of the positioning sensor and determining the positioning decision mode according to the GNSS signal strength includes:

[0014] Comparing the GNSS signal strength with a preset signal strength threshold, and determining a positioning decision mode based on the comparison result;

[0015] When the GNSS signal strength is greater than a preset signal strength threshold, the positioning decision mode is determined to be the GNSS-EKF fusion mode; when the GNSS signal strength is less than or equal to the preset signal strength threshold, the positioning decision mode is determined to be the SLAM+UWB particle filter fusion mode.

[0016] Furthermore, when it is determined that the positioning decision mode is the GNSS-EKF fusion mode, generating flight attitude control data according to the positioning decision mode includes:

[0017] Obtaining current location information based on GNSS positioning data, and converting the current location information into a local coordinate system position;

[0018] Obtain heading angle, pitch angle and roll angle based on IMU;

[0019] Predicting the flight state of the aircraft at a next moment and predicting the growth of uncertainty of the state based on the aircraft kinematic model and the heading angle, pitch angle, and roll angle;

[0020] A predicted measurement value is obtained according to the predicted flight state, the predicted measurement value is compared with the real-time GNSS observation value to obtain a real-time error, and flight attitude control data is generated according to the real-time error.

[0021] Furthermore, when the SLAM+UWB particle filter fusion mode is determined and the flight attitude control data is generated according to the positioning decision mode, the method includes:

[0022] Turn on the binocular vision, enable the UWB, initialize the ORB front end to start extracting feature points and building a sparse map;

[0023] SLAM extracts ORB feature points from the first frame and matches them between consecutive frames. It calculates the initial relative pose change based on the matched feature points and constructs a sparse point cloud map.

[0024] Generate a particle set near the initial estimated position, each particle in the particle set represents a hypothetical true position and is assigned an initial uniform weight;

[0025] Update the positions of all particles based on the movement of the flying robot described at the previous moment;

[0026] Collect the current UWB ranging data and calculate the distance of each particle to each UWB base station based on its position assumption;

[0027] Compare the predicted distance with the actual distance measurement, calculate the likelihood probability, update the particle weight, and perform resampling based on the particle weight distribution;

[0028] The center of the particle swarm after weighted average is used as the current estimated position of the aircraft, and flight attitude control data is generated according to the current estimated position and the relative posture change.

[0029] Furthermore, when performing feedforward compensation on the flight attitude according to the disturbance torque and the dynamic disturbance observer, the method includes:

[0030] Determining the disturbance torque according to the predicted wind speed data includes:

[0031] ;

[0032] ;

[0033] in, represents the disturbance torque, represents the fast convergence gain of sliding mode, Indicates the integral suppression gain; It represents the directional deviation between the prediction and the measurement. Indicates the IMU measured acceleration, represents the acceleration predicted from thrust and attitude;

[0034] The disturbance torque is converted into attitude torque compensation, and feedforward compensation is performed on the flight attitude according to the attitude torque compensation.

[0035] Furthermore, before starting and stopping the outer rotor of the flying robot based on Q-learning, the method further includes:

[0036] Collect rotor load rate, battery voltage and predicted wind speed data to establish the current state vector set;

[0037] Comparing the current state vector set with a historical control set, and controlling the start and stop of the outer rotor of the flying robot according to the comparison result; the historical control set includes a plurality of historical state vector sets and a plurality of historical recommended actions, and each historical state vector set corresponds to a historical recommended action;

[0038] When there is data in the historical control set whose similarity with the current state vector set is greater than or equal to a similarity threshold, a similarity set is established, and the start and stop of the outer rotor of the flying robot is controlled according to the historical recommended actions in the similarity set;

[0039] When the similarities between all historical state vector sets in the historical control set and the current state vector set are less than a similarity threshold, the start and stop of the outer rotor of the flying robot are controlled based on Q-learning.

[0040] Furthermore, when controlling the start and stop of the outer rotor of the flying robot based on Q-learning, the method includes:

[0041] Define the reward function R;

[0042] ;

[0043] Where R represents the reward function, Indicates the total power, Indicates the Euclidean distance of the aircraft from the hovering target point. represents the weight coefficient;

[0044] Initialize the Q table to 0; the Q table includes a state vector set s and an action a, and each combination of the state vector set s and the action a has a Q value;

[0045] Select a historical state vector set and choose an action a according to the greedy strategy;

[0046] After executing the action, observe the new state and immediate reward, and update the Q value. After training in the simulation environment for a preset number of times, a comparison Q table is obtained;

[0047] A recommended action is obtained based on the current state vector set and compared with the Q table, and the start and stop of the outer rotor of the flying robot is controlled according to the recommended action.

[0048] Furthermore, a similarity set is established, and the start and stop of the outer rotor of the flying robot is controlled according to the historical recommended actions in the similarity set, including:

[0049] Obtaining similarity between the current state vector set and each historical state vector set in the historical control set, and establishing historical state vector sets whose similarity is greater than or equal to the similarity threshold as the similarity set;

[0050] When the historical recommended actions corresponding to all the historical state vector sets in the similarity set are consistent, controlling the start and stop of the outer rotor of the flying robot according to the historical recommended actions;

[0051] When the historical recommended actions corresponding to all the historical state vector sets in the similarity set are inconsistent, the similarity score between each historical state vector in the similarity set and the current state vector is calculated, the weight is determined according to the similarity score, and the historical recommended actions corresponding to the historical state vector sets in the similarity set are weighted voted, and the start and stop of the outer rotor of the flying robot is controlled according to the historical recommended action with the largest total weight of votes.

[0052] Furthermore, when controlling the start and stop of the outer rotor of the flying robot according to the historical recommended action with the largest total vote weight, the process includes:

[0053] ;

[0054] in, Indicates the total weight of votes for the Ath historical recommended action, represents the similarity score between the current state vector set Sc and the history i-th state vector set, It represents the comparison result of the Ath historical recommended action and the ith historical recommended action, if and only if A=Ai, is 1, and n represents the total number of historical state vectors in the similarity set.

[0055] Compared with the existing technology, the beneficial effect of the present invention is that it ensures the complete and reliable functions by collecting and determining the connection status of module components such as positioning sensors, harmonic tilt-rotors, IMU, binocular vision, UWB and millimeter-wave radar before takeoff. Based on the positioning decision-making mechanism of GNSS signal strength, the GNSS-EKF fusion mode or SLAM+UWB particle filter fusion mode is intelligently switched, so that the flying robot can achieve continuous, stable and high-precision positioning in environments where GNSS is available or unavailable, thereby improving the robustness of hovering attitude perception. The rotor tilt angle is controlled by the harmonic reducer and servo motor closed-loop PID, eliminating the problems of traditional transmission slack and lag, achieving millisecond-level attitude response and high-precision tilt control, and enhancing the dynamic hovering stability of the aircraft. Combining the IMU and millimeter-wave radar to predict wind speed data, the real-time observation and feedforward compensation of the disturbance torque are used to effectively suppress the attitude fluctuations caused by gust interference, thereby improving the hovering accuracy of the aircraft under complex weather conditions. The central twin rotors provide a stable lift platform and incorporate intelligent start-stop control for the outer rotors based on a Q-learning algorithm, achieving load-adaptive energy consumption optimization, effectively extending flight endurance and improving system energy efficiency. Module self-checking ensures operational reliability, multi-modal positioning ensures environmental adaptability, high-speed tilt control and disturbance compensation guarantee wind-resistant hovering performance, and rotor management improves energy utilization. This ensures centimeter-level hovering accuracy while enhancing the aircraft's environmental adaptability, wind-resistant stability, and endurance.

[0056] On the other hand, the present application also provides an omnidirectional flying robot high-precision hovering system for applying the above-mentioned omnidirectional flying robot high-precision hovering method, comprising:

[0057] Module components, including positioning sensors, harmonic tiltrotors, IMUs, binocular vision, UWB, and millimeter-wave radars;

[0058] a collecting unit configured to collect connection status of module components in the flying robot and allow flying action when it is determined that the connection status of the module components is normal;

[0059] a judgment unit configured to, when the takeoff action is completed and it is determined to hover, collect the GNSS signal strength of the positioning sensor, determine a positioning decision mode based on the GNSS signal strength, and generate flight attitude control data based on the positioning decision mode, wherein the positioning decision mode includes a GNSS-EKF fusion mode and a SLAM+UWB particle filter fusion mode;

[0060] a processing unit configured to adjust the flight attitude of the flying robot according to the flight attitude control data and based on closed-loop PID control of a harmonic reducer and a servo motor;

[0061] a compensation unit configured to obtain predicted wind speed data based on the IMU and the millimeter-wave radar, determine a disturbance torque based on the predicted wind speed data, and perform feedforward compensation on the flight attitude based on the disturbance torque;

[0062] The adjustment unit is configured to control the continuous operation of the central dual rotors of the flying robot and control the start and stop of the middle and outer rotors of the flying robot based on Q-learning.

[0063] It is understandable that the above-mentioned high-precision hovering method and system for the omnidirectional flying robot have the same beneficial effects, which will not be described in detail here. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] Various other advantages and benefits will become apparent to those skilled in the art upon reading the detailed description of the preferred embodiment below. The accompanying drawings are for illustration purposes only and are not to be considered as limiting the present invention. The same reference symbols are used throughout the drawings to represent the same components. In the drawings:

[0065] Figure 1 A flowchart of a high-precision hovering method for an omnidirectional flying robot provided by an embodiment of the present invention;

[0066] Figure 2 Comparative curve of wind disturbance resistance experiment of high-precision hovering method for omnidirectional flying robot provided by the embodiment of the present invention;

[0067] Figure 3 This is the waveform of the harmonic drive response time test in the high-precision hovering method for an omnidirectional flying robot provided by an embodiment of the present invention;

[0068] Figure 4 A structural block diagram of a high-precision hovering system for an omnidirectional flying robot provided by an embodiment of the present invention;

[0069] Figure 5 This is a schematic structural diagram of an omnidirectional flying robot provided by an embodiment of the present invention.

[0070] Among them, 100, omnidirectional flying robot; 110, module assembly; 120, battery and main control cabin; 130, operating robotic arm; 140, central fixed rotor; 150, outer tilt rotor. DETAILED DESCRIPTION

[0071] Exemplary embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of the present disclosure are shown in the accompanying drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments set forth herein. On the contrary, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the present disclosure to those skilled in the art. It should be noted that, unless there is a conflict, the embodiments of the present disclosure and the features in the embodiments can be combined with each other. The present invention will be described in detail below with reference to the accompanying drawings and in conjunction with the embodiments.

[0072] In some embodiments of the present application, see Figure 1-3 As shown, a high-precision hovering method for an omnidirectional flying robot includes:

[0073] S100: Collect the connection status of the module components in the flying robot. When the connection status of the module components is determined to be normal, the flight action is allowed. The module components include positioning sensors, harmonic tilt rotors, IMU, binocular vision, UWB and millimeter wave radar.

[0074] S200: When the takeoff action is completed and it is determined to hover, the GNSS signal strength of the positioning sensor is collected, the positioning decision mode is determined according to the GNSS signal strength, and the flight attitude control data is generated according to the positioning decision mode. The positioning decision mode includes the GNSS-EKF fusion mode and the SLAM+UWB particle filter fusion mode.

[0075] S300: Adjust the flight attitude of the flying robot according to the flight attitude control data and based on the closed-loop PID control of the harmonic reducer and the servo motor.

[0076] S400: Obtain predicted wind speed data based on the IMU and millimeter-wave radar, determine the disturbance torque based on the predicted wind speed data, and perform feedforward compensation on the flight attitude based on the disturbance torque.

[0077] S500: Controls the continuous operation of the flying robot's central dual rotors and controls the start and stop of the flying robot's middle and outer rotors based on Q-learning.

[0078] Specifically, the connection status of the flying robot's modular components is collected. These modular components include the positioning sensor, harmonic tilt-rotor, IMU, binocular vision module, UWB positioning module, and millimeter-wave radar module. Based on the collected connection status data, a determination is made: flight maneuvers are permitted only when the connection status of all modular components is determined to be normal. After the takeoff maneuver is completed and a hovering mission is determined to be necessary, the GNSS signal strength of the positioning sensor is collected. Based on the GNSS signal strength, a positioning decision mode is dynamically determined. These modes include: a GNSS and extended Kalman filter (EKF) fusion mode when the GNSS signal is normal; and a binocular vision SLAM and UWB particle filter fusion mode when the GNSS signal is weak or unavailable. Based on the determined positioning decision mode, flight attitude control data is generated. Based on this flight attitude control data, the closed-loop PID control system constructed using a harmonic reducer and servo motor adjusts the flying robot's flight attitude. Leveraging the high reduction ratio and low hysteresis of the harmonic reducer and the fast response of the servo motor, high dynamic performance and high-precision attitude adjustment are achieved, improving the stability of the aircraft in dynamic environments. The predicted wind speed data is obtained through the IMU and millimeter-wave radar module, and the environmental disturbance torque is determined based on the predicted wind speed data. Based on the disturbance torque, the flight attitude control data is feedforward compensated. By introducing the disturbance prediction and feedforward compensation mechanism, the anti-disturbance capability of the aircraft in complex wind field environments is further enhanced, ensuring the stability and accuracy of hovering. In terms of thrust resource management, the dual rotors set in the center of the flying robot are controlled to operate continuously to maintain basic hovering thrust output, while the start and stop status of the outer rotors are controlled based on the Q-learning reinforcement learning method. According to the current load status of the aircraft, the degree of environmental disturbance and the remaining energy, the dynamic start and stop of the outer rotors are intelligently decided to optimize the allocation of overall thrust resources and energy consumption, effectively extending the flight time of the aircraft.

[0079] It is understandable that flight safety is improved by determining the connection status of module components, and the hovering accuracy and robustness in GNSS-restricted or failed environments are improved by dynamically switching positioning decision modes based on GNSS signal strength; high dynamic response and high-precision attitude adjustment are achieved through closed-loop PID control combining a harmonic reducer and a servo motor; wind disturbance prediction and disturbance torque feedforward compensation based on IMU and millimeter-wave radar effectively improve the stability of the aircraft in complex airflow environments; and the introduction of Q-learning reinforcement learning strategy for intelligent management of rotor start and stop improves the energy efficiency utilization of thrust resources and the endurance of the aircraft.

[0080] In some embodiments of the present application, collecting the GNSS signal strength of a positioning sensor and determining a positioning decision mode based on the GNSS signal strength includes comparing the GNSS signal strength with a preset signal strength threshold, and determining the positioning decision mode based on the comparison result. When the GNSS signal strength is greater than the preset signal strength threshold, the positioning decision mode is determined to be the GNSS-EKF fusion mode. When the GNSS signal strength is less than or equal to the preset signal strength threshold, the positioning decision mode is determined to be the SLAM+UWB particle filter fusion mode.

[0081] In some embodiments of the present application, when the positioning decision mode is determined to be the GNSS-EKF fusion mode, generating flight attitude control data according to the positioning decision mode includes:

[0082] The current position information is obtained based on the GNSS positioning data and converted into the local coordinate system position.

[0083] The heading angle, pitch angle and roll angle are obtained based on the IMU.

[0084] The flight state of the aircraft at the next moment is predicted based on the aircraft kinematic model and the heading angle, pitch angle and roll angle, and the uncertainty growth of the state is predicted.

[0085] The predicted measurement value is obtained according to the predicted flight state, the predicted measurement value is compared with the real-time GNSS observation value to obtain the real-time error, and the flight attitude control data is generated according to the real-time error.

[0086] Specifically, during the positioning mode selection phase, the GNSS signal strength measured by the positioning sensor is collected and compared with the preset signal strength threshold. The preset signal strength threshold here is set in advance based on the actual environmental needs and performance requirements, and can reasonably distinguish between the two situations of "GNSS reliable environment" and "GNSS weak / failed environment". When the GNSS signal strength is greater than the preset threshold, it means that the GNSS positioning in the current environment is reliable, and the GNSS-EKF fusion mode is determined to be adopted; if the GNSS signal strength is less than or equal to the preset threshold, it is determined that the GNSS is unreliable, and the SLAM+UWB particle filter fusion mode is switched to improve the positioning reliability in harsh environments.

[0087] Specifically, after adopting the GNSS-EKF fusion mode, the steps for generating flight attitude control data are as follows: The aircraft's current position (typically latitude, longitude, and altitude in a geographic coordinate system) is obtained using GNSS positioning data. This position information is then converted to a local navigation coordinate system (such as the NED or ENU coordinate system) for unified fusion with the aircraft's attitude and velocity data. The IMU (Inertial Measurement Unit) is used to obtain the aircraft's heading (yaw), pitch, and roll angles. These angles reflect the aircraft's spatial attitude changes and serve as the fundamental data source for state prediction and control. Based on the aircraft's kinematic model (such as a rigid-body dynamics model), the current heading, pitch, and roll angles are combined to predict the aircraft's motion state at the next moment, including position, velocity, and attitude. The uncertainty increase due to process noise is also predicted. This step is performed using the equation of state to ensure a reasonable estimate of the aircraft's state in the short term. Based on the predicted flight state, corresponding predicted measurements (such as the predicted position) are calculated and compared with real-time GNSS observations to determine the actual observation error. Based on the real-time error, the predicted state is updated through the extended Kalman filter (EKF), and the final flight attitude control data is output to guide the aircraft's attitude adjustment and position correction to ensure high-precision hovering control.

[0088] It is understandable that by setting the GNSS signal strength threshold to achieve dynamic positioning mode switching, the adaptability and reliability of the flying robot's positioning mode selection in complex environments are improved; when using the GNSS-EKF fusion mode, GNSS positioning and IMU attitude information are combined, and flight state prediction and real-time error correction are performed based on the kinematic model, which improves the accuracy and stability of the aircraft's posture estimation; the continuous prediction-observation-correction mechanism effectively suppresses the error accumulation caused by sensor noise and system dynamic changes, ensuring that the aircraft can maintain centimeter-level precise hovering even in signal fluctuations or short-term occlusion environments.

[0089] In some embodiments of the present application, when determining the SLAM+UWB particle filter fusion mode and generating flight attitude control data according to the positioning decision mode, it includes:

[0090] Turn on binocular vision, enable UWB, initialize the ORB front end to start extracting feature points and building a sparse map.

[0091] SLAM extracts ORB feature points from the first frame and matches them between consecutive frames. It calculates the preliminary relative pose changes based on the matched feature points and constructs a sparse point cloud map.

[0092] A particle set is generated near the initial estimated position. Each particle in the particle set represents a hypothetical true position and is assigned an initial uniform weight.

[0093] Update the positions of all particles based on the movement of the flying robot at the previous moment.

[0094] Collect the current UWB ranging data and calculate the distance of each particle to each UWB base station based on its position assumption.

[0095] Compare the predicted distance with the actual distance measurement, calculate the likelihood probability, update the particle weight, and perform resampling based on the particle weight distribution.

[0096] The center of the particle swarm after weighted average is used as the current estimated position of the aircraft, and the flight attitude control data is generated according to the current estimated position and relative posture changes.

[0097] Specifically, when the need to switch to SLAM+UWB particle filter fusion mode is detected, the binocular vision module (binocular camera) is immediately activated, and the UWB module is enabled for assisted positioning. Simultaneously, the ORB feature extraction and matching process is initialized, starting the front-end module of the ORB-SLAM system. Starting from the first frame, ORB feature points are extracted and matched between consecutive frames to infer the relative position changes (i.e., position and attitude changes) of the aircraft over short timescales. A sparse 3D map (sparse point cloud) is gradually constructed by accumulating the matched feature points. After the initial SLAM positioning is completed, a set of particles is generated near the initial estimated position. Each particle represents a possible true position hypothesis for the aircraft, and the particles are initially assigned the same weight (i.e., each particle is given equal credibility). Combined with the aircraft's motion information inferred from the IMU or attitude sensor (i.e., the predicted trajectory from the previous moment to the current moment), the state of all particles is predicted and updated, shifting the position of each particle based on the predicted motion. The current ranging data from the UWB module (the actual distance from the aircraft to several fixed base stations) is collected. For each particle in the particle set, the predicted distance from each UWB base station is calculated based on the particle's assumed position and compared with the actual UWB distance measurement. Based on the comparison error, the likelihood probability of each particle (i.e., the credibility of the particle's corresponding assumed position) is calculated, and the particle weight is updated based on the likelihood probability. To prevent particle degradation (i.e., the weights of most particles approach zero, resulting in a decrease in the effective number of particles), a resampling operation is performed based on the particle weight distribution: particles with low weights are discarded, and particles with high weights are retained or replicated, thereby maintaining the diversity of the particle set and positioning accuracy. The current optimal estimated position of the aircraft is determined by taking a weighted average of the positions of all particles; this is then combined with the continuous relative pose changes estimated during the SLAM process to generate the aircraft's flight attitude control data.

[0098] It is understandable that by integrating binocular vision SLAM and UWB particle filtering technology, not only can high-precision positioning be maintained in environments where GNSS signals fail or are blocked, but it can also effectively make up for the problem that single vision SLAM is prone to losing posture tracking in scenarios with sparse features and dynamic interference; particle filtering provides a multi-hypothesis position evaluation mechanism, enhances the adaptability to uncertain environments, and improves the robustness and reliability of aircraft positioning.

[0099] In some embodiments of the present application, when performing feedforward compensation on the flight attitude based on the disturbance torque and the dynamic disturbance observer, the method includes:

[0100] When determining the disturbance torque based on the predicted wind speed data, it includes:

[0101] ;

[0102] ;

[0103] in, represents the disturbance torque, represents the fast convergence gain of sliding mode, Indicates the integral suppression gain. It represents the directional deviation between the prediction and the measurement. Indicates the IMU measured acceleration, Represents the acceleration predicted from thrust and attitude.

[0104] The disturbance torque is converted into attitude torque compensation, and the flight attitude is feedforward compensated according to the attitude torque compensation.

[0105] Specifically, the system calculates predicted wind load effects based on ambient wind speed data collected by sensors such as millimeter-wave radar, combined with the aircraft's dimensions, center of mass position, and aerodynamic characteristics. Theoretical acceleration values ​​are inferred based on the aircraft's thrust output state (such as the current thrust magnitude and direction of each rotor) and attitude angle information (pitch angle, roll angle, etc.). Simultaneously, the system obtains actual measured acceleration data from the aircraft's IMU. By comparing the predicted acceleration with the measured acceleration, the deviation between the two is extracted, reflecting the actual impact of unmodeled external disturbances (such as wind disturbances) on the aircraft. During the calculation of the disturbance torque, a sliding mode fast convergence gain (to accelerate error convergence) and an integral suppression gain (to suppress low-frequency drift caused by accumulated errors) are introduced to improve the accuracy and response speed of the disturbance torque estimation. Combining this information, the disturbance torque—the torque affecting the aircraft's attitude due to external wind disturbances or uncertainties—is estimated in real time. This disturbance torque is converted into corresponding attitude torque compensation commands, and feedforward compensation is implemented in the flight control loop. In other words, before the attitude controller initiates corrective action due to attitude deviation, it proactively applies a compensating torque in the opposite direction based on the predicted disturbance to offset the impending attitude shift, thereby achieving rapid and precise attitude maintenance. This entire process, based on a coordinated mechanism of prediction, observation, and rapid compensation, minimizes the impact of environmental disturbances on the flying robot's stable hovering ability.

[0106] It can be understood that by quantifying the disturbance torque caused by the predicted wind speed, combining it with the dynamic observation of the IMU measured error, and using the sliding mode fast convergence and integral suppression mechanism to generate feedforward compensation, it is possible to actively apply compensation before the external disturbance significantly affects the flight attitude, thereby improving the wind resistance, response speed and attitude control accuracy of the flying robot; especially in complex environments (such as high-rise wind ducts and turbulent airflow areas under bridges), it can suppress attitude fluctuations caused by sudden wind disturbances, achieve more stable, reliable, and high-precision long-term hovering, and improve stability and safety in actual application scenarios.

[0107] In some embodiments of the present application, before controlling the start and stop of the outboard rotor of a flying robot based on Q-learning, the method further includes: collecting rotor load factor, battery voltage, and predicted wind speed data to establish a current state vector set. The current state vector set is compared with a historical control set, and the start and stop of the outboard rotor of the flying robot is controlled based on the comparison result. The historical control set includes several historical state vector sets and several historical recommended actions, and each historical state vector set corresponds to a historical recommended action.

[0108] Specifically, when there is data in the historical control set whose similarity to the current state vector set is greater than or equal to the similarity threshold, a similarity set is established, and the flying robot's mid-outboard rotors are controlled based on the historical recommended actions in the similarity set. When the similarity between all historical state vector sets in the historical control set and the current state vector set is less than the similarity threshold, the flying robot's mid-outboard rotors are controlled based on Q-learning.

[0109] In some embodiments of the present application, when controlling the start and stop of the outer rotor of a flying robot based on Q-learning, the process includes:

[0110] Define the reward function R.

[0111] ;

[0112] Where R represents the reward function, Indicates the total power, Indicates the Euclidean distance of the aircraft from the hovering target point. Represents the weight coefficient.

[0113] Initialize the Q table to 0. The Q table includes the state vector set s and the action a, and each combination of the state vector set s and the action a has a Q value.

[0114] Select a historical state vector set and choose an action a according to the greedy strategy.

[0115] After executing the action, observe the new state and immediate reward, and update the Q value. After training in the simulation environment for a preset number of times, a comparison Q table is obtained.

[0116] The recommended action is obtained based on the current state vector set and the comparison Q table, and the start and stop of the outer rotor of the flying robot are controlled according to the recommended action.

[0117] Specifically, before making rotor start / stop decisions, multi-dimensional information related to the flight state is collected to establish a current state vector set. This state vector set includes: the current load factor (ratio of thrust to maximum thrust) of each rotor; the current battery voltage (a measure of remaining energy); and the current predicted wind speed data (a measure of environmental disturbances). The current state vector set is compared with a pre-stored historical control set. This historical control set contains a large number of (historical state vector set → recommended action) pairs, namely, states collected in historical environments and corresponding rotor start / stop decisions. The similarity between the current state vector set and historical state vector sets is calculated. If data with a similarity greater than or equal to a set threshold exists, this data is extracted to form a similarity set. The current aircraft's rotor start / stop is directly controlled based on the corresponding historical recommended actions in the similarity set (such as turning on or off certain outer rotors), leveraging historical experience to accelerate decision-making. If the current state and historical data have low similarity (i.e., no human has encountered a similar situation), a Q-learning reinforcement learning strategy is initiated for autonomous exploration and decision-making.

[0118] In the Q-learning control part:

[0119] A reward function R is defined, which comprehensively considers: total power P (lower energy consumption is preferred); the offset e between the aircraft and the hovering target (smaller offset is preferred); w1 and w2 are weight coefficients that adjust the relative impact of energy consumption and positioning error. A Q-table (or Q-matrix) is initialized, listing all possible combinations of state vectors s and actions a. Each combination is assigned a Q-value, initially set to 0. Training is repeated in a simulation environment, using a greedy strategy (such as ε-greedy) to select actions, execute them, observe new states, obtain immediate rewards, and update Q-values. The optimal rotor start and stop strategy for different states is gradually learned. After training is complete, the final Q-table is obtained. During actual flight, the Q-table is quickly searched based on the current state vector, the optimal recommended action is selected, and rotor start and stop control is executed.

[0120] It is understandable that by combining rapid retrieval of historical experience with Q-learning reinforcement learning adaptive optimization, it is possible to use experience to respond quickly in known scenarios, and to continuously improve the quality of strategies through learning in unknown and complex environments, thereby realizing intelligent decision-making for the start and stop of the outer rotor of the flying robot; it not only effectively reduces energy consumption and improves flight endurance, but also maintains hovering accuracy and dynamic stability under different wind speed and load change conditions, thereby enhancing the autonomous adaptability and practical performance of the flying robot in complex and changing environments.

[0121] In some embodiments of the present application, a similarity set is established, and when controlling the start and stop of the outer rotor of a flying robot according to historical recommended actions in the similarity set, it includes: obtaining the similarity between the current state vector set and each historical state vector set in the historical control set, and establishing the historical state vector set whose similarity is greater than or equal to the similarity threshold as a similarity set.

[0122] Specifically, when the historical recommended actions corresponding to all historical state vector sets in the similarity set are consistent, the flying robot's mid-outer rotors are started and stopped according to the historical recommended actions. When the historical recommended actions corresponding to all historical state vector sets in the similarity set are inconsistent, the similarity score between each historical state vector in the similarity set and the current state vector is calculated. A weighted vote is then performed on the historical recommended actions corresponding to the historical state vector sets in the similarity set. The flying robot's mid-outer rotors are started and stopped according to the historical recommended action with the largest total vote weight.

[0123] In some embodiments of the present application, controlling the start and stop of the outer rotor of the flying robot according to the historical recommended action with the largest total vote weight includes:

[0124] ;

[0125] in, Indicates the total weight of votes for the Ath historical recommended action, represents the similarity score between the current state vector set Sc and the history i-th state vector set, It represents the comparison result of the Ath historical recommended action and the ith historical recommended action, if and only if A=Ai, is 1, and n represents the total number of historical state vectors in the similarity set.

[0126] Specifically, the current state vector set (including rotor load rate, battery voltage, predicted wind speed, etc.) is collected, and the similarity is calculated with each historical state vector set in the historical control set to obtain the similarity score between the current state and the historical state. All historical state vector sets with similarities greater than or equal to the preset similarity threshold are screened out to form a similarity set. The similarity set represents a batch of historical experience data that is closest to the current flight state characteristics. If the corresponding historical recommended actions in the similarity set are completely consistent (that is, all historical experiences make the same decision in this environment, such as uniformly recommending "outer rotor closure"), then this historical recommended action is directly adopted to quickly make a start-stop control decision. If there is inconsistency (that is, the recommended actions given in different historical states are different), a weighted voting mechanism is used to determine the optimal action. Calculate the similarity score between the current state vector set and the historical state vector set in each similarity set. For each possible recommended action A, count the sum of the similarity scores of all historical state vectors consistent with action A to obtain the total weight of votes for action A. Among them, It represents the comparison result of the Ath historical recommended action and the ith historical recommended action, if and only if A=Ai, is 1, otherwise it is 0. The historical recommended action with the largest total vote weight is selected as the execution action of the current rotor start and stop control.

[0127] It is understandable that through the voting mechanism based on similarity weighting, when there are differences in historical recommended actions, the degree of match between each historical sample and the current state can be comprehensively considered to make rotor start and stop decisions that are more in line with the actual environmental needs; compared with traditional simple majority voting or blind reinforcement learning exploration, it has the advantages of faster response, more stable decision-making, and higher tolerance to abnormal data, thereby further improving the energy efficiency management level and flight attitude stability of the flying robot in complex dynamic environments, effectively extending the flight endurance and enhancing the adaptability.

[0128] In the above embodiment, the connection status of module components such as positioning sensors, harmonic tilt-rotors, IMU, binocular vision, UWB and millimeter-wave radar is collected and determined before takeoff to ensure complete and reliable functions. Based on the positioning decision mechanism of GNSS signal strength, the GNSS-EKF fusion mode or SLAM+UWB particle filter fusion mode is intelligently switched, so that the flying robot can achieve continuous, stable and high-precision positioning in environments where GNSS is available or unavailable, thereby improving the robustness of hovering attitude perception. The rotor tilt angle is controlled by the harmonic reducer and servo motor closed-loop PID, eliminating the problems of traditional transmission slack and lag, achieving millisecond-level attitude response and high-precision tilt control, and enhancing the dynamic hovering stability of the aircraft. In combination with the IMU and millimeter-wave radar predicted wind speed data, the disturbance torque is observed in real time and feedforward compensation is used to effectively suppress the attitude fluctuations caused by gust interference, thereby improving the hovering accuracy of the aircraft in complex weather conditions. The central twin rotors provide a stable lift platform and incorporate intelligent start-stop control for the outer rotors based on a Q-learning algorithm, achieving load-adaptive energy consumption optimization, effectively extending flight endurance and improving system energy efficiency. Module self-checking ensures operational reliability, multi-modal positioning ensures environmental adaptability, high-speed tilt control and disturbance compensation guarantee wind-resistant hovering performance, and rotor management improves energy utilization. This ensures centimeter-level hovering accuracy while enhancing the aircraft's environmental adaptability, wind-resistant stability, and endurance.

[0129] In another preferred embodiment based on the above embodiment, refer to Figure 4 As shown, this embodiment provides an omnidirectional flying robot high-precision hovering system for applying the above-mentioned omnidirectional flying robot high-precision hovering method, including:

[0130] Module components include positioning sensors, harmonic tiltrotors, IMU, binocular vision, UWB and millimeter wave radar.

[0131] The collecting unit is configured to collect the connection status of the module components in the flying robot, and allows the flying action when it is determined that the connection status of the module components is normal.

[0132] The judgment unit is configured to collect the GNSS signal strength of the positioning sensor when the takeoff action is completed and it is determined to hover, determine the positioning decision mode according to the GNSS signal strength, and generate flight attitude control data according to the positioning decision mode. The positioning decision mode includes the GNSS-EKF fusion mode and the SLAM+UWB particle filter fusion mode.

[0133] The processing unit is configured to adjust the flight attitude of the flying robot according to the flight attitude control data and based on the closed-loop PID control of the harmonic reducer and the servo motor.

[0134] The compensation unit is configured to obtain predicted wind speed data based on the IMU and the millimeter-wave radar, determine the disturbance torque based on the predicted wind speed data, and perform feedforward compensation on the flight attitude based on the disturbance torque.

[0135] The adjustment unit is configured to control the continuous operation of the central dual rotors of the flying robot and to control the start and stop of the middle and outer rotors of the flying robot based on Q-learning.

[0136] Specifically, see Figure 5 As shown, the omnidirectional flying robot 100 employing the above-described method or system includes a module assembly 110, a battery and main control cabin 120, an operating manipulator arm 130, a central fixed rotor 140, and outboard tilt rotors 150. The battery and main control cabin 120 includes a data acquisition unit, a judgment unit, a processing unit, a compensation unit, and a compensation unit. A cup-shaped harmonic reducer (model CSF-17-100, reduction ratio 1:100) is directly connected to a brushless motor (T-Motor MN5208, peak torque 0.5 Nm) and the rotor shaft, which is then fixed to the end of the hexagonal arm via a flange. The hexagonal fuselage reduces the crosswind drag coefficient to 0.12 (compared to 0.25 for a conventional quadrotor).

[0137] It is understood that by collecting and determining the connection status of modules such as the positioning sensor, harmonic tilt-rotor, IMU, binocular vision, UWB, and millimeter-wave radar before takeoff, the complete and reliable functionality is ensured. A positioning decision mechanism based on GNSS signal strength intelligently switches between GNSS-EKF fusion mode and SLAM+UWB particle filter fusion mode, enabling the flying robot to achieve continuous, stable, high-precision positioning in both GNSS-available and unavailable environments, improving the robustness of hovering attitude perception. The rotor tilt angle is controlled by a harmonic reducer and servo motor closed-loop PID, eliminating the slack and lag problems of traditional transmissions, achieving millisecond-level attitude response and high-precision tilt control, and enhancing the dynamic hovering stability of the aircraft. Combining wind speed data predicted by the IMU and millimeter-wave radar, real-time observation of disturbance torque and feedforward compensation are used to effectively suppress attitude fluctuations caused by gusts of wind, improving the aircraft's hovering accuracy in complex weather conditions. The central twin rotors provide a stable lift platform and incorporate intelligent start-stop control for the outer rotors based on a Q-learning algorithm, achieving load-adaptive energy consumption optimization, effectively extending flight endurance and improving system energy efficiency. Module self-checking ensures operational reliability, multi-modal positioning ensures environmental adaptability, high-speed tilt control and disturbance compensation guarantee wind-resistant hovering performance, and rotor management improves energy utilization. This ensures centimeter-level hovering accuracy while enhancing the aircraft's environmental adaptability, wind-resistant stability, and endurance.

[0138] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware. Furthermore, the present application may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0139] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems) and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0140] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0141] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0142] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the field should understand that the specific implementation methods of the present invention can still be modified or replaced by equivalents. Any modification or equivalent replacement that does not depart from the spirit and scope of the present invention should be covered by the scope of protection of the claims of the present invention.

Claims

1. A high-precision hovering method for an omnidirectional flying robot, characterized in that: include: Collecting the connection status of module components in the flying robot, and allowing flight actions when the module components are determined to be connected normally, the module components including positioning sensors, harmonic tilt rotors, IMUs, binocular vision, UWB, and millimeter wave radars; When the takeoff action is completed and it is determined to hover, the GNSS signal strength of the positioning sensor is collected, a positioning decision mode is determined according to the GNSS signal strength, and flight attitude control data is generated according to the positioning decision mode. The positioning decision mode includes a GNSS-EKF fusion mode and a SLAM+UWB particle filter fusion mode. Adjusting the flight attitude of the flying robot according to the flight attitude control data and based on closed-loop PID control of a harmonic reducer and a servo motor; Obtaining predicted wind speed data according to the IMU and the millimeter-wave radar, determining a disturbance torque according to the predicted wind speed data, and performing feedforward compensation on the flight attitude according to the disturbance torque; Control the continuous operation of the central dual rotors of the flying robot, and control the start and stop of the outer rotors of the flying robot based on Q-learning; When feedforward compensation is performed on the flight attitude according to the disturbance torque and the dynamic disturbance observer, the method includes: Determining the disturbance torque according to the predicted wind speed data includes: ； ； in, represents the disturbance torque, represents the fast convergence gain of sliding mode, Indicates the integral suppression gain; It represents the directional deviation between the prediction and the measurement. Indicates the IMU measured acceleration, represents the acceleration predicted from thrust and attitude; Converting the disturbance torque into attitude torque compensation, and performing feedforward compensation on the flight attitude according to the attitude torque compensation; Before controlling the start and stop of the outer rotor of the flying robot based on Q-learning, the method further includes: Collect rotor load rate, battery voltage and predicted wind speed data to establish the current state vector set; Comparing the current state vector set with a historical control set, and controlling the start and stop of the outer rotor of the flying robot according to the comparison result; the historical control set includes a plurality of historical state vector sets and a plurality of historical recommended actions, and each historical state vector set corresponds to a historical recommended action; When there is data in the historical control set whose similarity with the current state vector set is greater than or equal to a similarity threshold, a similarity set is established, and the start and stop of the outer rotor of the flying robot is controlled according to the historical recommended actions in the similarity set; When the similarities between all historical state vector sets in the historical control set and the current state vector set are less than a similarity threshold, controlling the start and stop of the outer rotor of the flying robot based on Q-learning; Establishing a similarity set, and controlling the start and stop of the outer rotor of the flying robot according to the historical recommended actions in the similarity set, including: Obtaining similarity between the current state vector set and each historical state vector set in the historical control set, and establishing historical state vector sets whose similarity is greater than or equal to the similarity threshold as the similarity set; When the historical recommended actions corresponding to all the historical state vector sets in the similarity set are consistent, controlling the start and stop of the outer rotor of the flying robot according to the historical recommended actions; When the historical recommended actions corresponding to all the historical state vector sets in the similarity set are inconsistent, the similarity score between each historical state vector in the similarity set and the current state vector is calculated, the weight is determined according to the similarity score, and the historical recommended actions corresponding to the historical state vector sets in the similarity set are weighted voted, and the start and stop of the outer rotor of the flying robot is controlled according to the historical recommended action with the largest total weight of votes.

2. The high-precision hovering method for an omnidirectional flying robot according to claim 1, characterized in that: Collecting the GNSS signal strength of the positioning sensor and determining the positioning decision mode according to the GNSS signal strength includes: Comparing the GNSS signal strength with a preset signal strength threshold, and determining a positioning decision mode based on the comparison result; When the GNSS signal strength is greater than a preset signal strength threshold, the positioning decision mode is determined to be the GNSS-EKF fusion mode; when the GNSS signal strength is less than or equal to the preset signal strength threshold, the positioning decision mode is determined to be the SLAM+UWB particle filter fusion mode.

3. The high-precision hovering method for an omnidirectional flying robot according to claim 2, characterized in that: When it is determined that the positioning decision mode is the GNSS-EKF fusion mode, generating flight attitude control data according to the positioning decision mode includes: Obtaining current location information based on GNSS positioning data, and converting the current location information into a local coordinate system position; Obtain heading angle, pitch angle and roll angle based on IMU; Predicting the flight state of the aircraft at a next moment and predicting the growth of uncertainty of the state based on the aircraft kinematic model and the heading angle, pitch angle, and roll angle; A predicted measurement value is obtained according to the predicted flight state, the predicted measurement value is compared with the real-time GNSS observation value to obtain a real-time error, and flight attitude control data is generated according to the real-time error.

4. The high-precision hovering method for an omnidirectional flying robot according to claim 2, characterized in that: When the SLAM+UWB particle filter fusion mode is determined and the flight attitude control data is generated according to the positioning decision mode, the method includes: Turn on the binocular vision, enable the UWB, initialize the ORB front end to start extracting feature points and building a sparse map; SLAM extracts ORB feature points from the first frame and matches them between consecutive frames. It calculates the initial relative pose change based on the matched feature points and constructs a sparse point cloud map. Generate a particle set near the initial estimated position, each particle in the particle set represents a hypothetical true position and is assigned an initial uniform weight; Update the positions of all particles based on the movement of the flying robot described at the previous moment; Collect the current UWB ranging data and calculate the distance of each particle to each UWB base station based on its position assumption; Compare the predicted distance with the actual distance measurement, calculate the likelihood probability, update the particle weight, and perform resampling based on the particle weight distribution; The center of the particle swarm after weighted average is used as the current estimated position of the aircraft, and flight attitude control data is generated according to the current estimated position and the relative posture change.

5. The high-precision hovering method for an omnidirectional flying robot according to claim 1, characterized in that: When controlling the start and stop of the outer rotor of the flying robot based on Q-learning, the method includes: Define the reward function R; ； Where R represents the reward function, Indicates the total power, Indicates the Euclidean distance of the aircraft from the hovering target point. represents the weight coefficient; Initialize the Q table to 0; the Q table includes a state vector set s and an action a, and each combination of the state vector set s and the action a has a Q value; Select a historical state vector set and choose an action a according to the greedy strategy; After executing the action, observe the new state and immediate reward, and update the Q value. After training in the simulation environment for a preset number of times, a comparison Q table is obtained; A recommended action is obtained based on the current state vector set and compared with the Q table, and the start and stop of the outer rotor of the flying robot is controlled according to the recommended action.

6. The high-precision hovering method for an omnidirectional flying robot according to claim 1, characterized in that: Controlling the start and stop of the outer rotor of the flying robot according to the historical recommended action with the largest total vote weight includes: ； in, Indicates the total weight of votes for the Ath historical recommended action, represents the similarity score between the current state vector set Sc and the history i-th state vector set, It represents the comparison result of the Ath historical recommended action and the ith historical recommended action, if and only if A=Ai, is 1, and n represents the total number of historical state vectors in the similarity set.

7. A high-precision hovering system for an omnidirectional flying robot, for applying the high-precision hovering method for an omnidirectional flying robot according to any one of claims 1 to 6, characterized in that: include: Module components, including positioning sensors, harmonic tiltrotors, IMUs, binocular vision, UWB, and millimeter-wave radars; a collecting unit configured to collect connection status of module components in the flying robot and allow flying action when it is determined that the connection status of the module components is normal; a judgment unit configured to, when the takeoff action is completed and it is determined to hover, collect the GNSS signal strength of the positioning sensor, determine a positioning decision mode based on the GNSS signal strength, and generate flight attitude control data based on the positioning decision mode, wherein the positioning decision mode includes a GNSS-EKF fusion mode and a SLAM+UWB particle filter fusion mode; a processing unit configured to adjust the flight attitude of the flying robot according to the flight attitude control data and based on closed-loop PID control of a harmonic reducer and a servo motor; a compensation unit configured to obtain predicted wind speed data based on the IMU and the millimeter-wave radar, determine a disturbance torque based on the predicted wind speed data, and perform feedforward compensation on the flight attitude based on the disturbance torque; The adjustment unit is configured to control the continuous operation of the central dual rotors of the flying robot and control the start and stop of the middle and outer rotors of the flying robot based on Q-learning.

Images