Autonomous Exploration Methods for UAVs in Indoor Environments Based on Navigation Experience Knowledge

Abstract

This invention discloses an autonomous exploration method for unmanned aerial vehicles (UAVs) in indoor environments based on navigation experience knowledge. The method includes: Step 1, multi-sensor data acquisition and state estimation; Step 2, constructing an indoor 3D map; Step 3, generating exploration targets; Step 4, experience-driven local waypoint search; Step 5, path smoothing and flight trajectory optimization; and Step 6, closed-loop autonomous exploration execution. This invention can effectively improve the exploration efficiency, flight safety, and environmental mapping integrity of UAVs in unknown indoor environments.

Description

Technical Field

[0001] This invention belongs to the field of robot autonomous navigation and intelligent unmanned systems technology, specifically relating to a navigation method for autonomous exploration and flight path generation of small rotary-wing unmanned aerial vehicles in unknown indoor environments. Background Technology

[0002] Small rotary-wing unmanned aerial vehicles (UAVs) are widely used in disaster search and rescue, industrial inspection, underground space exploration, underground parking garage inspection, and laboratory patrols due to their high maneuverability, hovering capability, flexible deployment, and ability to traverse ground obstacles. In these scenarios, the environment typically lacks prior maps, and because they are located indoors, underground, or semi-enclosed spaces, they often cannot receive signals from global navigation satellite systems. Therefore, UAVs need to rely on onboard sensors to autonomously complete environmental perception, target discovery decisions, and flight path planning. Improving exploration efficiency and flight safety in unknown indoor environments under conditions of limited computing power and payload has become a key issue in UAV autonomous exploration research.

[0003] During autonomous exploration in unknown indoor environments, the upper-level exploration task typically needs to continuously output the target location based on the currently acquired environmental information, while the local planning module is responsible for generating a flight path between the UAV's current position and the target position. Because environmental information is constantly updated, the local planning module needs to balance computational real-time performance, obstacle avoidance safety, and integration with the flight control system.

[0004] Existing path planning techniques typically include search-based methods, sampling-based methods, and methods that combine learning and optimization. For example, graph search and random sampling expansion can be performed on 3D grid maps, or smoothing can be applied after obtaining discrete paths. These methods can generate paths from the current location to the target point to a certain extent. However, for autonomous exploration missions of small UAVs in unknown indoor environments, performing high-density searches or large-scale sampling directly in the 3D environment incurs high computational costs, which is detrimental to real-time airborne operation. Generating only discrete polyline paths often results in sharp corners, discontinuous speeds, abrupt acceleration changes, and dynamic infeasibility, making them difficult to use directly for flight control. Furthermore, if the path generation process does not fully utilize obstacle distance information, the trajectory may become too close to obstacles, resulting in insufficient flight safety margins.

[0005] In summary, the existing technology has at least the following shortcomings: 1. Existing local path planning methods have high computational overhead in complex three-dimensional environments, making it difficult to meet the real-time requirements of airborne operations; 2. The discrete waypoint generation process lacks unified handling of local navigation experience constraints such as safety distance, turning continuity, and unpassable backtracking, and there is still room for improvement in the safety margin and stability of local paths; 3. Existing discrete paths often lack sufficient smoothness and dynamic feasibility, making it difficult to directly meet the requirements for safe flight control of UAVs.

[0006] Therefore, there is an urgent need to propose a UAV navigation method that can incorporate navigation experience knowledge into the local flight path generation process and combine it with trajectory optimization to achieve autonomous exploration of unknown indoor environments. Summary of the Invention

[0007] Therefore, the technical problem to be solved by the present invention is that existing indoor UAVs have insufficient robustness in positioning and mapping during autonomous exploration, large computational load in local three-dimensional path search during exploration, and discrete paths are difficult to meet the requirements of obstacle avoidance, smooth flight and dynamic feasibility of the aircraft.

[0008] To address the aforementioned technical problems, this invention proposes an autonomous exploration method for unmanned aerial vehicles (UAVs) in indoor environments based on navigation experience knowledge, comprising the following steps: Step 1: Multi-sensor data acquisition and state estimation Real-time synchronous acquisition of binocular or depth image data, inertial measurement unit data, magnetometer data, barometer data and flight control feedback data; initial pose calculation of UAV is completed by visual-inertial tight coupling optimization; attitude angle, barometric altitude and magnetic heading data are then fused by loose coupling filtering to output UAV state variables including three-dimensional position, three-axis velocity, body attitude angle and yaw angle. Step 2: Construct an indoor 3D map Based on the real-time pose of the UAV, the depth image is back-projected and a spatial point cloud is generated. A sliding window-style 3D occupied grid map centered on the UAV is constructed. The grid idle / occupied state is updated using a ray casting algorithm, and the grid occupancy probability is updated by combining a Bayesian algorithm. Simultaneously, a Euclidean symbolic distance field map is constructed to complete obstacle safety expansion processing. Step 3: Exploring Target Generation Based on the updated environmental information, the exploration target generation module determines the current target location corresponding to the autonomous exploration task and sends the target location to the local waypoint search module; Step 4: Experience-driven local waypoint search For the current position of the UAV and the target exploration viewpoint, a local three-dimensional obstacle tensor is constructed. The three-dimensional path search is decomposed into multiple orthogonal two-dimensional sub-programming problems through dimensional decomposition. The value iteration network is used to complete the two-dimensional space action value inference. The multi-dimensional action value is integrated to output the three-dimensional space movement decision and iteratively generate a collision-free discrete flight waypoint sequence. Step 5: Path Smoothing and Flight Trajectory Optimization Adaptive redundancy pruning is performed on discrete flight waypoints, and segmented flight time allocation is completed based on flight distance and maximum UAV flight speed. Polynomial interpolation is used to generate an initial continuous path, which is then transformed into a uniform B-spline trajectory. A multi-objective optimization function integrating smoothness, obstacle avoidance, dynamics, boundary constraints, and time constraints is constructed, and the L-BFGS algorithm is used to optimize the trajectory control points and time parameters, outputting the final flight trajectory that meets the UAV flight control requirements. Step Six: Closed-Loop Autonomous Exploration and Execution The optimized final flight trajectory is sent to the UAV flight controller to drive the UAV to complete autonomous flight. During the flight, steps one to six are continuously repeated to update the environmental map and exploration decisions in real time until the proportion of unknown areas in the indoor environment is lower than the preset threshold or there are no effective boundary units to be explored, thus completing the autonomous exploration operation in the entire environment.

[0009] In the above technical solution, in step one, when the tracking quality of the visual sensor is lower than a preset threshold, the existing map coordinate system is locked, and the fusion weight of the inertial measurement unit and the flight controller sensor data is increased; after the visual tracking state returns to normal, the conventional multi-sensor fusion positioning mode is returned to normal.

[0010] In the above technical solution, in step two, for each newly entered depth data frame, the system first performs depth validity detection and outlier filtering, and then projects the depth pixels into spatial points according to the in-camera pose.

[0011] In the above technical solution, in step three, the exploration target generation module outputs the current target position based on the current environmental information and the autonomous exploration task status, and the current target position can be adjusted as the environmental information is updated.

[0012] In the above technical solution, in step four, the value iteration network relies on the Bellman optimal equation to complete the spatial value propagation deduction. In the process of generating discrete waypoints, collision detection, maximum flight step size constraints and flight back-off mechanisms are added to improve the obstacle avoidance reliability of local path search.

[0013] In the above technical solution, in step five, the trajectory optimization process takes the minimum integral value of flight acceleration as the optimization objective, constructs obstacle repulsion and obstacle avoidance costs based on the Euclidean symbolic distance field map, and utilizes the characteristics of the derivative of uniform B-spline trajectory to transform the dynamic constraints of the UAV's maximum speed and maximum acceleration into geometric constraints of trajectory control points.

[0014] In the above technical solution, in step six, if the UAV experiences abnormal positioning, path planning failure, or trajectory optimization non-convergence during flight, the flight protection mode is immediately triggered, the active exploration flight is suspended, and the autonomous planning process is restarted after a new round of environmental map updates.

[0015] The advantages of this invention are: 1. This invention addresses the problem of high computational cost in local 3D pathfinding within complex indoor environments by proposing a waypoint search method based on dimensionality decomposition and value iteration networks. This method transforms the highly complex 3D pathfinding problem into multiple low-dimensional sub-problems for reasoning, and obtains locally feasible waypoints through fusion decision-making. Compared to methods that directly perform high-overhead 3D searches, this invention reduces computational burden while maintaining local obstacle avoidance guidance capabilities, improves the real-time performance of local planning, and is more suitable for small rotary-wing UAV platforms with limited onboard computing power.

[0016] 2. After generating local waypoints, this invention further combines adaptive path pruning, polynomial interpolation, and uniform B-spline nonlinear optimization to perform continuous and smoothing processing on the discrete path, and integrates safety distance constraints and aircraft dynamics constraints into the trajectory optimization process. Therefore, this invention can output a flight trajectory that simultaneously meets obstacle avoidance requirements, smoothness requirements, and dynamic executability requirements, avoiding the problems of abrupt corner changes, trajectory discontinuities, and difficulty in direct execution found in traditional polyline paths, thereby improving the stability and safety of UAV flight.

[0017] 3. This invention constructs a hierarchical autonomous exploration framework suitable for indoor unknown environments by organically combining an exploration target generation module, experience-driven local pathfinding, and trajectory optimization. This framework effectively reduces redundant movement and decision-making delays in complex topological scenes, obstacle-dense scenes, and scenes with limited local visibility, improving overall exploration efficiency and enhancing the stability and engineering feasibility of UAV autonomous exploration. Simulation and real-world verification have shown that it achieves higher exploration efficiency in scenarios such as large mazes, multi-column open spaces, underground parking garages, and laboratories. The method outperforms mainstream comparative methods in terms of exploration time and flight trajectory length, demonstrating that this invention has significant practical application value in autonomous exploration tasks in indoor unknown environments. Attached Figure Description

[0018] The present invention will be further described below with reference to the accompanying drawings and embodiments: Figure 1 This is a top view of the drone of the present invention.

[0019] Figure 2 This is a schematic diagram of the main hardware framework of the present invention.

[0020] Figure 3 This is a schematic diagram of the main software framework of the present invention.

[0021] Figure 4 This is a simplified flowchart of the unmanned aerial vehicle (UAV) indoor exploration system of the present invention. Detailed Implementation

[0022] This invention provides an autonomous exploration method for unmanned aerial vehicles (UAVs) in indoor environments based on navigation experience knowledge. It is applicable to a small rotary-wing UAV platform 1 equipped with a visual sensor, inertial measurement unit, flight controller, and onboard computing unit. The UAV platform 1 includes: a depth camera 2, an inertial measurement unit, a flight controller, a magnetometer, a barometer, an onboard computing unit 3, and a flight actuator 4. The top view of the UAV, a schematic diagram of the main hardware framework, and a schematic diagram of the main software framework are shown below. Figures 1 to 3 As shown.

[0023] See Figure 4 As shown, the present invention provides an autonomous exploration method for unmanned aerial vehicles (UAVs) in indoor environments based on navigation experience knowledge, comprising the following steps: Step 1: Multi-sensor data acquisition and state estimation This invention acquires binocular or depth images, inertial measurement unit data, magnetometer data, and barometer data in real time. It employs a visual-inertial fusion positioning method to obtain the UAV's pose.

[0024] After takeoff, the onboard computing unit receives image data, depth data, inertial measurement unit data, magnetometer data, barometer data, and flight controller feedback data in real time, and caches and synchronizes them according to timestamps. The image sampling frequency is 15 Hz, the inertial measurement unit sampling frequency is 100 Hz, and the magnetometer and barometer sampling frequencies are 20 to 50 Hz.

[0025] In implementation, the camera and inertial measurement unit are first calibrated for extrinsic parameters and time synchronization. Then, a positioning and mapping method based on visual-inertial tight coupling optimization is used to jointly optimize image features, inertial pre-integral quantities, and historical keyframes to obtain initial estimates of the UAV's current position, velocity, and attitude. Next, the attitude angles, barometric altitude, and magnetic heading output by the flight controller are fused and corrected using loosely coupled filtering to obtain the final state variables used in the planning module. These final state variables include at least three-dimensional position, three-axis velocity, body attitude angles, and yaw angle.

[0026] To improve stability during long-term indoor operation, when visual tracking quality falls below a threshold, the system maintains the map coordinate system from the previous moment and increases the weights of inertial and flight control sensors; once image tracking is restored, it re-enters the normal fusion state. This approach ensures continuous and stable pose inputs required for subsequent map building, candidate viewpoint evaluation, and trajectory planning.

[0027] The purpose of this step is to provide a stable and reliable position and attitude basis for subsequent environmental mapping, viewpoint assessment, and trajectory planning.

[0028] Step 2: Construct an indoor 3D map The onboard computing unit backprojects the depth image onto the world coordinate system based on the current pose estimation results to generate a point cloud, and then writes the point cloud into a local voxel map. The local map is represented in a 3D occupancy grid format, and the local map update range is set to a sliding window area centered on the UAV, based on the size of the UAV's flight space.

[0029] For each newly entered depth data frame, the system first performs depth validity detection and outlier filtering, and then projects the depth pixels into spatial points according to the poses of the participants in the camera. Subsequently, the system uses a ray casting method to update the voxels between the sensor and the spatial points to the idle state, and updates the voxels near the landing point to the occupied state. Finally, the system uses Bayesian update or log-probability update to maintain the grid occupancy probability.

[0030] Based on the occupied grid map, a Euclidean symbolic distance field map is further constructed to obtain the distance and gradient values ​​from any location to the nearest obstacle. Then, the obstacles are expanded according to the sum of the UAV's envelope radius and the safety margin to form a safe occupied area. The distance and gradient values ​​are repeatedly used in subsequent candidate viewpoint safety judgments, discrete waypoint selection, and trajectory optimization.

[0031] Step 3: Exploring Target Generation Based on the updated environmental information, the exploration target generation module determines the current target location corresponding to the autonomous exploration task and sends the current target location to the local path generation module.

[0032] Step 4: Experience-driven local waypoint search Based on the current UAV pose and the target viewpoint selected in step three, an obstacle tensor is constructed in the local 3D environment map. The 3D search problem is decomposed into multiple orthogonal 2D subproblems through dimensionality decomposition. Obstacle feature maps and target distance field feature maps are extracted for each 2D subproblem and input into the value iteration network. This allows the network to explicitly simulate the value propagation process and output the action value of each 2D slice. MDP is usually represented as a quintuple. .in, and These are the state space and the action space, respectively. The reward function quantifies the actions performed in a given state. And transfer to state Instant rewards obtained , It is a state transition model, given a state and actions In this case, the system transitions to state probability , This is a discount factor used to weigh long-term returns against immediate rewards. Within the MDP framework, the strategy... A probabilistic mapping from state to action is defined.

[0033] value function Indicates the state Follow the strategy The recursive property of the expected cumulative return that can be obtained is defined by the Bellman equation: Traditional value iteration is a classic dynamic programming algorithm used to solve the optimal strategy in a multiplication table (MDP). And its corresponding optimal value function. Its core is based on iterative updates of the Bellman optimal equation until convergence: Subsequently, the system fuses the action values ​​of multiple two-dimensional sub-problems according to spatial direction to obtain a comprehensive score of multiple candidate actions in the three-dimensional neighborhood, and selects the action with the highest comprehensive score as the current movement direction. After the UAV moves one step forward according to this action, the above reasoning process is repeated until it reaches the target viewpoint neighborhood, the search step limit is reached, or the local map determines that the path ahead is impassable, thereby generating a discrete waypoint sequence.

[0034] The action values ​​of multiple two-dimensional sub-problems are fused to obtain the probability distribution or priority of multiple candidate actions in the three-dimensional neighborhood; then, based on the action distribution, the next way point is inferred step by step from the current UAV position, and a discrete way point sequence leading to the target viewpoint is iteratively generated.

[0035] During the generation process, local collision detection, maximum step size constraints, and backtracking mechanisms can be combined to delete collision actions and reselect suboptimal actions to improve the robustness of local pathfinding.

[0036] Step 5: Path Smoothing and Flight Trajectory Optimization Adaptive path pruning is performed on the discrete waypoints output in step four, retaining necessary topological nodes; time allocation is performed based on the geometric distance between waypoints, for the first... Segment Path Interval Allocate execution time It satisfies the following formula: Among them, fixed constant , Let be the Euclidean distance between the two path points. This is the maximum flight speed of the drone.

[0037] An initial continuous path is generated using polynomial interpolation, and the overall trajectory is denoted as . , No. Segment trajectory Should be within the original path interval , As shown in the formula below: in, , is the first one to be solved Duan Di The coefficient vector of a polynomial of order 1.

[0038] To ensure the geometric connectivity and dynamic feasibility of the trajectory, the trajectory smoothing problem is modeled as a constrained optimization problem. The sum of the squared integrals of the jerk of all trajectory segments is taken as the objective function: in, For the first Segment trajectory The third derivative of , i.e., jerk.

[0039] The initial continuous path is then converted into a uniform B-spline trajectory representation, one B-order spline curve Depend on Control points and time node vector Uniquely determined. When the interval between adjacent time points is constant, i.e. When the B-spline curve is a uniform B-spline, it can be represented as a linear combination of control points: in, It is a B-spline basis function defined by the Cox-deBoor recursive formula.

[0040] A nonlinear objective function is constructed, including smoothness cost, obstacle avoidance cost, dynamic feasibility cost, boundary constraint cost, and time cost. By iteratively optimizing the control point position and time parameters, the final flight trajectory that satisfies the safety distance constraint and the dynamic constraints such as the speed and acceleration of the UAV is obtained.

[0041] Among them, the obstacle avoidance cost term is directly constructed using the obstacle distance values ​​and gradient information provided by the Euclidean symbolic distance field map; the dynamic feasibility constraint utilizes the derivative closure property of the uniform B-spline trajectory to transform the velocity and acceleration constraints into a control point optimization problem.

[0042] By adjusting the control points Spacing between nodes To optimize the generated trajectory: in, Smoothing costs Obstacle avoidance costs , cost of dynamic feasibility Boundary constraint cost and time cost The weighting coefficients.

[0043] For smoothing cost The jerk is approximated using the third-order finite difference of the control points of the uniform B-spline: Regarding the cost of obstacle avoidance According to the convex hull property of uniform B-splines, when each control point If all points are within the safe zone, then the trajectory will be confined within the convex hull formed by the control points: Among them, control points Distance to the nearest obstacle and gradient Provided by a real-time maintained ESDF map. A quadratic penalty is applied to generate a repulsive force that pushes the control point away from the obstacle: in, This is the safe distance threshold.

[0044] Cost of kinetic feasibility To ensure that the trajectory does not exceed the drone's maximum speed and maximum acceleration By utilizing the closure property of the derivative of a uniform B-spline, higher-order dynamic constraints are transformed into geometric constraints between adjacent control points. The relationship between velocity control points and position control points is as follows: in, For the speed control point, construct the speed penalty term: in, Indicates speed control point The magnitudes of the components along each of the x, y, and z axes. Similarly, the relationship between the acceleration control point and the position control point is: in, Let the acceleration control point be used. Construct the acceleration penalty term: in, Indicates acceleration control point The magnitude of the components along the x, y, and z axes. The cost of dynamic feasibility. Equal to the sum of the two: Cost of boundary constraints The trajectory must strictly satisfy dynamic constraints at both ends. Boundary cost constraints at the trajectory start point. for: in, These represent the position, velocity, and acceleration at the start of the trajectory, respectively.

[0045] Similarly, boundary cost constraints at the end of the trajectory for: in, These represent the position, velocity, and acceleration at the end of the trajectory, respectively. Boundary constraint cost. Equal to the sum of the two: Regarding the cost of time The total time for a third-order uniform B-spline trajectory is: Finally, the L-BFGS algorithm is used to jointly optimize the control point position and time parameters, and the output is the final trajectory that can be directly sent to the flight controller.

[0046] In summary, adaptive path pruning is performed on the discrete waypoint sequence obtained in step four; when the straight-line passability condition is met between adjacent discontinuous waypoints, intermediate redundant waypoints are deleted, and only necessary topological turning points are retained.

[0047] Based on the distance between adjacent waypoints after pruning and the maximum speed of the UAV, execution time is allocated to each path segment; based on the time allocation result, piecewise polynomial interpolation is used to generate an initial continuous path.

[0048] The L-BFGS optimization algorithm is used to jointly solve the B-spline control points and time parameters, and output the final flight trajectory that satisfies the safety distance constraint and dynamic constraint. When the optimization fails to converge or the trajectory has a collision risk, it can be rolled back to re-prune, reallocate time, or regenerate the local waypoint sequence.

[0049] Step Six: Closed-Loop Autonomous Exploration and Execution The final flight trajectory output from step five is sent to the flight controller, which then drives the drone to fly using the position control and attitude control modules.

[0050] During flight, steps one through six are repeated until no unexplored boundary clusters meet the conditions on the map, the proportion of unknown areas is lower than a preset threshold, or the environmental coverage reaches a preset value, thus completing the autonomous exploration of the entire indoor unknown environment.

[0051] The present invention will be further described in detail below with reference to specific embodiments. It should be understood that these embodiments are only for illustrating the present invention and are not intended to limit the scope of protection of the present invention. Equivalent substitutions or modifications made by those skilled in the art without departing from the concept of the present invention should fall within the scope of protection of the present invention.

[0052] Complete system deployment example: This embodiment uses a small rotary-wing UAV equipped with a depth camera as the execution platform. The UAV body includes a frame, motors, propellers, flight controller, battery, onboard computing unit, depth camera, inertial measurement unit, magnetometer, and barometer. The depth camera is fixedly mounted at the front of the fuselage, while the inertial measurement unit and flight controller are fixedly mounted near the center of gravity of the fuselage. The onboard computing unit communicates with the sensors and flight controller via a serial port, USB interface, or Ethernet port.

[0053] After the system powers on, it first completes time synchronization of all sensors, external loading, and map initialization. Then, it starts the state estimation thread, map update thread, target generation thread, local pathfinding thread, and trajectory optimization thread. The state estimation thread outputs the UAV's real-time pose; the map update thread updates the occupied grid map and Euclidean symbolic distance field map based on pose and depth data; the target generation thread provides the current target position; the local pathfinding thread generates discrete waypoints; and the trajectory optimization thread outputs the final trajectory. The flight controller receives the trajectory and drives the UAV to fly. These modules run continuously in a loop until the exploration ends.

[0054] A basic implementation example of autonomous indoor drone exploration based on navigation experience knowledge: This embodiment uses a small rotary-wing UAV equipped with a depth camera as the execution platform. The UAV's onboard components include a depth camera, an inertial measurement unit, a flight controller, a magnetometer, a barometer, and an onboard computing unit. After takeoff, the UAV first acquires image, depth, inertial, and flight controller sensor data, and obtains the current pose estimation result. Subsequently, a local 3D environment map is constructed based on the depth image and the current pose, and an Euclidean symbolic distance field map is established on the basis of the grid map to provide an environmental representation basis for subsequent exploration decisions and trajectory optimization.

[0055] After acquiring status and environmental information, the upper-level target generation module outputs the current target location based on the current environmental information and task execution status. This example does not limit the specific method of generating the current target location; the current target location is used to guide the local path generation module to generate the flight path within the current execution cycle.

[0056] In the local path generation stage, the system extracts local 3D obstacle tensors based on the current UAV position and the current target viewpoint, and uses a dimensionality decomposition method to break down the 3D local pathfinding problem into multiple 2D sub-problems. For each 2D sub-problem, obstacle feature maps and target guidance feature maps are generated and input into a value iteration network for value propagation and action reasoning. Then, the action values ​​on multiple 2D planes are combined to obtain the probability distribution of candidate actions in 3D space, and a discrete waypoint sequence from the current position to the target viewpoint is gradually generated.

[0057] During the trajectory generation phase, adaptive path pruning is performed on the discrete waypoint sequence to retain necessary topological nodes. Time allocation is performed based on the distance between nodes, and an initial continuous path is generated using polynomial interpolation. The initial path is then converted into a uniform B-spline trajectory, and a nonlinear optimization model is constructed, incorporating smoothness cost, obstacle avoidance cost, dynamic feasibility cost, boundary constraint cost, and time cost. The control points and time parameters are jointly optimized to obtain the final flight trajectory that satisfies obstacle avoidance constraints and UAV dynamic constraints. Finally, the flight control module controls the UAV to fly to the target viewpoint along the optimized trajectory and repeats the above process until there are no more unexplored boundary clusters on the map, thus completing the autonomous exploration of the unknown indoor environment.

[0058] Example of local experience-based pathfinding and trajectory optimization: After the current target viewpoint is output in Implementation Example 3, the system extracts local obstacle tensors around the current position of the UAV and maps this three-dimensional space into multiple two-dimensional programming sub-problems. For each two-dimensional sub-problem, an obstacle feature map and a target guidance feature map are generated, and then input into a value iteration network to obtain the action value distribution on the two-dimensional plane. The system fuses multiple two-dimensional value distributions to obtain the three-dimensional candidate action scores at the current time, and selects the action with the highest score that meets the safety constraints to gradually form a discrete waypoint sequence.

[0059] After discrete waypoints are generated, the system performs path pruning: if a waypoint has an unobstructed direct connection to its subsequent waypoints, redundant intermediate waypoints are deleted, retaining only critical waypoints necessary for obstacle avoidance and turning. Then, time allocation is performed based on the distances between critical waypoints to generate an initial continuous trajectory, which is then converted into a uniform B-spline trajectory. Nonlinear optimization is used to adjust the control points, ensuring the optimized trajectory maintains overall smoothness while satisfying obstacle safety distance constraints and dynamic constraints such as UAV speed and acceleration. The resulting final trajectory can be directly used for flight control execution, avoiding problems such as abrupt corner changes, dynamic non-executability, and insufficient safety margins inherent in discrete polyline paths.

[0060] Closed-loop autonomous exploration example: This embodiment illustrates the overall closed-loop workflow. After the UAV takes off, the system first initializes the map and state estimation; then it enters a loop: collecting sensor data, updating environmental information, determining the target location, performing local pathfinding, optimizing the trajectory, and executing the flight. After each target location visit is completed, the system re-executes the above process based on the newly perceived results.

[0061] When the upper-level exploration task module determines that the autonomous exploration task has been completed, the system controls the UAV to hover or return to base. If an abnormal state estimation occurs during flight, a local map becomes impassable, or trajectory optimization fails, the system enters protection mode, stops moving forward, and waits for the next map update before replanning. This closed-loop mechanism ensures the continuous autonomous operation capability of this invention in unknown indoor environments.

Claims

1. A method for autonomous exploration of unmanned aerial vehicles (UAVs) in indoor environments based on navigation experience knowledge, characterized in that: Includes the following steps: Step 1: Multi-sensor data acquisition and state estimation Real-time synchronous acquisition of binocular or depth image data, inertial measurement unit data, magnetometer data, barometer data and flight control feedback data; initial pose calculation of UAV is completed by visual-inertial tight coupling optimization; attitude angle, barometric altitude and magnetic heading data are then fused by loose coupling filtering to output UAV state variables including three-dimensional position, three-axis velocity, body attitude angle and yaw angle. Step 2: Construct an indoor 3D map Based on the real-time pose of the UAV, the depth image is back-projected and a spatial point cloud is generated. A sliding window-style 3D occupied grid map centered on the UAV is constructed. The grid idle / occupied state is updated using a ray casting algorithm, and the grid occupancy probability is updated by combining a Bayesian algorithm. Simultaneously, a Euclidean symbolic distance field map is constructed to complete obstacle safety expansion processing. Step 3: Exploring Target Generation Based on the updated environmental information, the exploration target generation module determines the current target location corresponding to the autonomous exploration task and sends the target location to the local waypoint search module; Step 4: Experience-driven local waypoint search For the current position of the UAV and the target exploration viewpoint, a local three-dimensional obstacle tensor is constructed. The three-dimensional path search is decomposed into multiple orthogonal two-dimensional sub-programming problems through dimensional decomposition. The value iteration network is used to complete the two-dimensional space action value inference. The multi-dimensional action value is integrated to output the three-dimensional space movement decision and iteratively generate a collision-free discrete flight waypoint sequence. Step 5: Path Smoothing and Flight Trajectory Optimization Adaptive redundancy pruning is performed on discrete flight waypoints, and segmented flight time allocation is completed based on flight distance and maximum UAV flight speed. Polynomial interpolation is used to generate an initial continuous path, which is then transformed into a uniform B-spline trajectory. A multi-objective optimization function integrating smoothness, obstacle avoidance, dynamics, boundary constraints, and time constraints is constructed, and the L-BFGS algorithm is used to optimize the trajectory control points and time parameters, outputting the final flight trajectory that meets the UAV flight control requirements. Step Six: Closed-Loop Autonomous Exploration and Execution The optimized final flight trajectory is sent to the UAV flight controller to drive the UAV to complete autonomous flight. During the flight, steps one to six are continuously repeated to update the environmental map and exploration decisions in real time until the proportion of unknown areas in the indoor environment is lower than the preset threshold or there are no effective boundary units to be explored, thus completing the autonomous exploration operation in the entire environment.

2. The method for autonomous exploration of unmanned aerial vehicles (UAVs) in indoor environments based on navigation experience knowledge as described in claim 1, characterized in that, In step one, when the tracking quality of the visual sensor is lower than a preset threshold, the existing map coordinate system is locked and the fusion weight of the inertial measurement unit and the flight controller sensor data is increased. Once the visual tracking status returns to normal, the system will revert to the regular multi-sensor fusion positioning mode.

3. The method for autonomous exploration of unmanned aerial vehicles (UAVs) in indoor environments based on navigation experience knowledge as described in claim 1, characterized in that, In step two, for each newly entered depth data frame, the system first performs depth validity detection and outlier filtering, and then projects the depth pixels into spatial points based on the in-camera pose.

4. The method for autonomous exploration of unmanned aerial vehicles in indoor environments based on navigation experience knowledge as described in claim 1, characterized in that, In step three, the exploration target generation module outputs the current target location based on the current environmental information and the autonomous exploration task status. The current target location can be adjusted as the environmental information is updated.

5. The method for autonomous exploration of unmanned aerial vehicles in indoor environments based on navigation experience knowledge as described in claim 1, characterized in that, In step four, the value iteration network relies on the Bellman optimal equation to complete the spatial value propagation deduction. In the process of generating discrete waypoints, collision detection, maximum flight step size constraints and flight back-off mechanisms are added to improve the obstacle avoidance reliability of local path search.

6. The method for autonomous exploration of unmanned aerial vehicles in indoor environments based on navigation experience knowledge according to claim 1, characterized in that, In step five, the trajectory optimization process takes minimizing the integral value of flight acceleration as the optimization objective, constructs obstacle repulsion and obstacle avoidance costs based on the Euclidean symbolic distance field map, and utilizes the derivative characteristics of uniform B-spline trajectory to transform the dynamic constraints of the UAV's maximum speed and maximum acceleration into geometric constraints of the trajectory control points.

7. The method for autonomous exploration of unmanned aerial vehicles in indoor environments based on navigation experience knowledge as described in claim 1, characterized in that, In step six, if the UAV experiences an abnormal positioning status, path planning failure, or trajectory optimization non-convergence during flight, the flight protection mode is immediately triggered, suspending active exploration flight and waiting for a new round of environmental map updates before restarting the autonomous planning process.

Images