Due to their high flexibility, low cost, and ease of handling, Unmanned Aerial Vehicles (UAVs) are often used to perform difficult tasks in complex environments. Stable and reliable path planning capability is the fundamental demand for UAVs to accomplish their flight tasks. Most researches on UAV path planning are carried out under the premise of known environmental information, and it is difficult to safely reach the target position in the face of unknown environment. Thus, an autonomous collision-free path planning algorithm for UAVs in unknown complex environments (APPA-3D) is proposed. An anti-collision control strategy is designed using the UAV collision safety envelope, which relies on the UAV's environmental awareness capability to continuously interact with external environmental information. A dynamic reward function of reinforcement learning combined with the actual flight environment is designed and an optimized reinforcement learning action exploration strategy based on the action selection probability is proposed. Then, an improved RL algorithm is used to simulate the UAV flight process in unknown environment, and the algorithm is trained by interacting with the environment, which finally realizes autonomous collision-free path planning for UAVs. The comparative experimental results in the same environment show that APPA-3D can effectively guide the UAV to plan a safe and collision-free path from the starting point to the target point in an unknown complex 3D environment.
Unmanned Aerial Vehicles (UAVs) are widely used in a variety of scenarios due to their abilities of high flexibility, high productivity, ease of maneuverability, and adapting to hazardous environments. The increasing complexity of flight environments requires UAVs to have the ability to interact with highly dynamic and strongly real-time space operating environments, which put forward new demands for UAVs' autonomy and safety. UAVs detect and determine whether there is a potential conflict in the future period through the sensors so that they can maintain a certain safe distance from the dynamic/static obstacles in the airspace, and thus plan an ideal flight path from the starting point to the target point and avoid conflicts.
Unlike civil aircraft, UAVs usually perform tasks in lower airspace. There are many static obstacles in lower airspace such as buildings, trees, and dynamic aircraft. Flight conflict is a state when the distance between two aircraft in the direction of horizontal, longitudinal, or vertical is less than a specific interval resulting in the aircraft being at risk[
Graph: Figure 1Schematic diagram of UAV perception and avoidance.
For UAVs, the ability of SAA is extremely important. The ability of path planning in the avoidance function of UAVs is an important foundation for the basis for them to complete the flight task. Complex flight environments put forward higher demands for path planning algorithms of UAVs, thus the research in autonomous obstacle avoidance path planning algorithms for UAVs is necessary.
It has to find the optimal flight path from the initial location to the target location under the constraints of environmental factors such as terrain, weather, threats, and flight performance of autonomous path planning for UAVs. Significantly, the optimal path does not always mean the shortest path or a straight line between two locations; instead, the UAV aims to find a safe path under limited power and flight task. There are a lot of UAV path planning algorithms, such as the Voronoi diagram algorithm, Rapidly-exploring Random Tree (RRT) algorithm, A* algorithm, etc. However, these algorithms cannot deal with dynamic environments effectively because they require global environmental information to calculate the optimal result. Once the environment changes, the original results will fail. Furthermore, the process of recalculating the optimal results is too slow for real-time operations because of the large number of calculations required. The above algorithms may still be effective if the obstacle is moving slowly. But when moving faster, the movement of the surrounding vehicles may cannot be predicted thus result in a collision. These shortcomings limit the application of the above algorithms to UAVs in real, dynamic environments.
To address these shortcomings, reinforcement learning algorithms are applied to the path planning process. Reinforcement learning (RL) is a branch of machine learning. UAVs can learn through continuous interaction with the environment, using training and learning to master the environment gradually, and optimize the state-behavior continuously to obtain the optimal strategy through the feedback (rewards) given by the environment, which is closer to the human learning process.
Compared with traditional algorithms, RL performs better when the environment is unknown and dynamic. Moreover, the inference speed and generalization of RL have advantages in real-time decision-making tasks. Therefore, the path planning algorithm based on RL has certain advantages in solving the UAV path planning problem in unknown and dynamic environments.
This paper considers the real-time and location limitation characteristics of path planning and refers to the existing research on UAV path planning problems and collision avoidance strategies for various stationary/motion threats. An autonomous collision-free path planning algorithm for UAVs in unknown complex 3D environments (APPA-3D) is proposed. Thus, UAVs can perform tasks with APPA-3D more safely and efficiently in complex flight environments. Firstly, the UAV spherical safety envelope is designed to research the anti-collision avoidance strategy, which will be used as an action plan for UAVs to realize dynamic obstacle avoidance. Secondly, we assume that the environment model when path planning is unknown, so the UAV needs to have the ability to learn and adjust flight state intelligently according to its surroundings. In this paper, the traditional model-free RL algorithm is improved to reduce the complexity of the algorithm and adapt to the demands of UAV path planning in an unknown complex 3D environment. It takes into account the search efficiency while guaranteeing the optimal search path.
Compared with the existing research, the innovative work of this paper mainly manifests in the following several aspects:
Based on the UAV environment sensing capability, a collision safety envelope is designed, and the anti-collision control strategy is studied concerning the Near Mid Air Collision (NMAC) rules for civil airliners and the International Regulations for Preventing Collisions at Sea (COLREGS). It provides a theoretical basis for UAVs to carry out collision detection and avoidance schemes, which can detect and avoid dynamic threats effectively in the flight environment.
To address the difficulty of convergence of traditional algorithms in solving 3D path planning. The artificial potential field method (APF) is used to optimize the mechanism of reward function generation in RL. The optimized algorithm can output the dynamic reward function by combining the actual flight environment information. Thus, the problems of path planning convergence difficulty, unreachable target point and model stop learning in high dimensional space caused by sparse reward function are solved.
Aiming at the "exploration-exploitation dilemma" of RL in the path planning process of UAVs, an RL action exploration strategy based on action selection probability is proposed. The strategy dynamically adjusts the action selection strategy by combining the size of the value function in different states, thus solving the RL exploration-exploitation problem and improving the efficiency of path search.
The rest of the paper is organized as follows: Section "Related research on UAV path planning" introduces the research status of UAV path planning; an anti-collision control strategy for UAVs is designed in Section "UAV anti-collision control strategy"; an Autonomous collision-free path planning algorithm is proposed in Section "Design of autonomous collision-free path planning algorithm for UAVs"; simulation experiment design and result analysis are presented in Section "Experiment and results"; the paper is summarized in Section "Conclusion".
Autonomous mobile robots (AMRs) has attracted more and more attention due to their practicality and potential uses in the modern world[
UAV path planning refers to the formulation of the optimal flight path from the initial location to the target location, considering environmental factors such as terrain, meteorology, threats, and their flight performance constraints The aim is to improve the reliability and safety of UAVs while ensuring the efficiency of their task execution.
A lot of research has been done on the UAV path planning problem. Sampling-based path planning algorithms are widely used in UAV path planning due to their simplicity, intuitiveness, and ease of implementation. A simple sampling-based path planning algorithm is the Voronoi diagram algorithm[
In node-based path planning algorithms, Dijkstra's algorithm searches for the shortest path by cyclic traversal[
Computational intelligence (CI) algorithms can provide solutions to NP-hard problems with many variables. CI algorithms are a group of nature-inspired methods, which have been raised as a solution for these problems. They can address complex real-world scenarios that algorithms. Genetic Algorithm (GA)[
Real-time and autonomy in complex flight environments are important indicators for measuring different path-planning algorithms. In the above algorithms, the sampling-based path planning algorithms reduce the traversal search space by sampling, sacrificing the optimality of paths in exchange for a shorter computation time. As the size of the environment increases, the number of operation iterations increases dramatically, making it difficult to achieve simultaneous optimization accuracy and optimal paths in 3D complex environments. Node-based path planning algorithms can obtain the optimal path between the start and endpoints. However, as the environment expands, the dimensionality increases and the number of search nodes increases, the computational size of these algorithms will increase dramatically. Intelligent biomimetic algorithms optimize paths in a mutation-like manner, which can better handle unstructured constraints in complex scenarios. However, its variational solving process requires a long iteration period and cannot be adapted to path planning in dynamic environments[
In response to the limitations of the traditional algorithms mentioned above, a new feasible solution is to update the distance information between the UAVs and the obstacles and target points in real-time and feed it back to the UAV, as well as to make real-time adjustments to its flight status and maneuvers[
In the face of the problem that the environment model is unknown and the transfer probability and value function are difficult to determine, the RL algorithm of interactive learning with the environment to obtain the optimal strategy needs to be a model-free RL algorithm. The Q-Learning (QL) algorithm is one of the most commonly used model-free RL algorithms and has been widely applied to solve path-planning problems. Reference[
All in all, RL takes rewards from exploring the environment as training data by imitating the learning process of human beings and trains itself without requiring preset training data. The path planning algorithm of UAV based on RL senses the state information of obstacles in the environment continuously and inputs the information into the algorithm, The optimal collision-free path can be obtained by adjusting the flight state of the UAV through RL, which can solve the problems of poor real-time and long planning time of traditional trajectory planning.
However, in practice, due to the complexity of the flight environment, the traditional RL algorithms do not run well in complex scenarios. More concretely, the memory size of a Q-table increases exponentially as the dimensionality of the state space or action space associated with the environment increases[
UAVs are unable to obtain complete priori information about the environment during the flight, and can only obtain the information within a certain range centered on themselves through various onboard sensors such as Light Detection And Ranging (LiDAR), and vision sensors. The maximum distance that the sensors of a UAV can detect is defined as
Graph: Figure 2UAV spherical safety envelope profile.
Before applying APPA-3D to solve the UAV path planning problem, an anti-collision avoidance strategy should be designed. The purpose is to adjust the UAV's flight status such as direction or speed in response to dynamic obstacles (such as other vehicles in the airspace, birds, etc.) to achieve obstacle avoidance. The design of the anti-collision avoidance strategy refers to the method of setting up collision zones in (NMAC) and (COLREGS).
When a dynamic obstacle enters a UAV's MZ, a collision avoidance strategy will be triggered to reduce the risk of collision until the distance between the UAV and the obstacle is greater than
Graph: Figure 3The relative position of the UAV and the dynamic obstacle.
In Fig. 3, when the flight path of the dynamic obstacle is in the same straight line as the UAV, the two are about to have an opposing conflict or pursuit conflict, and their relative positions are schematically shown in Fig. 3A,B. The vertical direction vector
When the flight path of the dynamic obstacle is in the same straight line as the UAV, the two flight paths cross and conflict, and their relative positions are shown in Fig. 3C. Vector
Based on the different collision scenarios generated by the relative positions of the UAV and dynamic obstacles, four corresponding anti-collision avoidance strategies are designed, as shown in Fig. 4.
Graph: Figure 4Schematic of UAV anti-collision avoidance strategies.
In Fig. 4A, there is a risk of opposing conflict between the dynamic obstacle and the UAV. Similar to the method shown in Fig. 3A, the UAV will fly along the merging vector
In Fig. 4B,C, there is a risk of cross-conflict between dynamic obstacles and UAVs. Similar to the method shown in Fig. 3C, the UAV will fly along the merging vector direction
In Fig. 4D, dynamic obstacles are in the UAV's path of travel and moving at a speed less than the UAV. There is a risk of pursuit and conflict between the dynamic obstacles and the UAV. Similar to the method shown in Fig. 3B, the UAV will fly along the merging vector direction
The basic framework of RL is shown in Fig. 5.
Graph: Figure 5Framework of RL.
Essentially, RL is the use of the Agent to interact with the environment constantly, and obtain the optimal value function
Graph
2
Graph
Thus, the problem of finding the optimal strategy translates into finding the largest of the action state value functions produced under all strategies.
RL-based path planning algorithm allows UAVs to learn and gain rewards through constant interaction with the surroundings through trial and error with little knowledge of the environment and is, therefore, suitable for UAV 's path planning under complex conditions. The advantage of an RL-based path planning algorithm is that it can realize path planning in the absence of a priori information about the environment and is highly searchable, but it suffers from the problem of reward sparsity[
APPA-3D first combines the principle of the APF method and designs an adaptive reward function. Dynamic rewards are generated in real time by judging the effectiveness of UAV movements with environmental information. Secondly, to address the "exploration-utilization dilemma" of RL in the UAV path planning process, an RL action exploration strategy based on action selection probability is proposed. The strategy dynamically adjusts the action selection strategy by combining the size of the value function in different states, to solve the exploration-utilization problem of RL and improve the efficiency of path search.
The basic idea of path planning with the APF[
The current position of the UAV is denoted as
3
Graph
In Eq. (
4
Graph
Define the repulsive potential field function as Eq. (
5
Graph
In Eq. (
6
Graph
7
Graph
Thus the net force
8
Graph
The reward function is used to evaluate the actions of the Agent. In traditional RL algorithms, the Agent can only obtain the positive and negative sparse reward function by reaching the target point or colliding with an obstacle. The model does not receive any feedback until it receives the first reward, which may cause the model to stop learning and fail to improve. This reward function will make the algorithm convergence difficult, and in most states cannot reflect the good or bad of its action choice. he sparse reward function
9
Graph
In Eq. (
To solve the difficult problem of sparse rewards, in this paper, combined with the artificial potential field algorithm, the gravitational force generated by the target point and the repulsive force generated by the obstacle on the agent are converted into the reward or punishment obtained by the agent after performing the action
10
Graph
In Eq. (
The Euclidean distance between the starting point of the agent and the target is denoted by
11
Graph
12
Graph
The hyperbolic tangent function can map all will any real number to (− 1, 1). The hyperbolic tangent function can be written as Eq. (
13
Graph
As shown in Fig. 6A, when the obstacle is in the SZ, or no obstacle is detected, the UAV is only affected by the gravitational force
14
Graph
Graph: Figure 6Direction of UAV movement when the obstacle in different zones.
In Eq. (
From Eq. (
As shown in Fig. 6B, when the obstacle is in the CZ, the UAV is affected by the repulsive force
15
Graph
where:
From Eq. (
As shown in Fig. 6C, when the obstacle is within the MZ, the risk of UAV collision with the obstacle is high. To prevent conflicts, A collision avoidance strategy
16
Graph
The adaptive reward function is consistent with RL. By converting the reward values of each action-state into continuous value between (− 1, 1), the problem of sparse reward functions is solved. The adaptive reward function solves the problem that traditional reward functions can only earn positive or negative rewards by reaching a target point or colliding with an obstacle, and other actions do not receive any positive or negative feedback. The adaptive reward functions are generated by determining the validity of the executed action and environmental information. Compared to the traditional sparse reward function, the adaptive reward function proposed in this paper combines the good performance of APF to make the reward accumulation process smoother, and can also reflect the relationship between the current state of the UAV and the target state.
In the process of constant interaction with the environment, the Agent keeps exploring different states and obtains feedback on different actions. Exploration helps the Agent to obtain feedback through continuous experimentation, and Exploitation is where the Agent refers to the use of existing feedback to choose the best action.
On the one hand, RL obtains more information by exploring more of the unknown action space to search for the global optimal solution, but a large amount of exploration reduces the performance of the algorithm and leads to the phenomenon of non-convergence of the algorithm. On the other hand, too much exploitation will fail to choose the optimal behavior because of the unknown knowledge of the environment. Therefore how to balance exploration and utilization is an important issue for the Agent to continuously learn in interaction.
There is a contradiction between "exploration" and "exploitation ", as the number of attempts is limited, and strengthening one naturally weakens the other. Excessive exploration of the unknown action space can degrade the performance of the algorithm and lead to non-convergence of the algorithm while obtaining more information to search for a globally optimal solution. In contrast, too much exploitation prevents the selection of optimal behavior because of the unknown knowledge of the environment. This is the Exploration—Exploitation dilemma faced by RL. To maximize the accumulation of rewards, a better compromise between exploration and exploitation must be reached.
Action exploration strategies can be categorized into directed and undirected exploration methods. The directed exploration approach reduces the blindness in the pre-exploration phase of action exploration and thus improves the exploration efficiency by introducing a priori knowledge into the action exploration strategy. directed exploration methods, on the other hand, make a compromise between exploration and exploitation by setting parameters, and the usual approaches are the
The
17
Graph
In Eq. (
The Softmax distribution strategy makes a tradeoff between exploration and exploitation based on the average reward of currently known actions. If the average rewards of the maneuvers are comparable, the probability of selecting each maneuver is also comparable; if the average reward of some maneuvers is significantly higher than that of other maneuvers, the probability of their selection is also significantly higher.
The action assignment for the Softmax distribution strategy is based on the Boltzmann distribution, which is represented by Eq. (
18
Graph
In Eq. (
Both
Action selection probability represents the probability value that an Agent chooses to perform an action in a given state. As shown in Eq. (
19
Graph
In Eq. (
The action selection probability is dynamically adjusted as the size of the value function of the action changes. During the RL process, Agent in state
The Agent updates the action selection probability after completing an action, according to the size of the state action value function. The update rule is as Eq. (
20
Graph
In Eq. (
In the initial phase of the algorithm, each action has the same probability of being selected by the Agent, i.e., the action selection probabilities
After an Agent completes the exploration of an action, if this exploration results in
The pseudo-code for APPA-3D is shown in Algorithm 1:
Graph: Algorithm 1APPA-3D.
To verify the feasibility of an autonomous path planning algorithm in complex 3D environments (APPA-3D) for UAVs, this paper selects real environment maps to conduct simulation experiments. The UAV's range of action is limited to the map, and if the UAV moves outside the range of the map or above the low altitude limit altitude, it is determined that a collision has occurred. The starting point for UAV path planning is represented by a black dot, and the target point is represented by a red dot. The maximum flight altitude is 1 km above the peak line, and the no-fly zone is indicated by a green cylinder. The UAV needs to avoid mountains and no-fly zones to fly from the starting point to the target point.
The anti-collision avoidance strategies experiments were designed to verify whether APPA-3D can achieve anti-collision avoidance strategies while implementing path planning. Figures 7, 8 and 9 are simulation experimental diagrams of anti-collision avoidance strategies for UAVs.
Graph: Figure 7Opposing Conflict Avoidance Simulation.
Graph: Figure 8Cross-conflict avoidance simulation.
Graph: Figure 9Pursuit conflict avoidance simulation.
As shown in Fig. 7, in the opposing conflict avoidance simulation, the intrusion direction of the dynamic obstacle is set to be directly opposite the movement direction of the UAV. When a dynamic obstacle is detected, the UAV chooses to turn right to avoid the dynamic obstacle according to the anti-collision avoidance strategy. The reward function is recalculated after finishing the anti-collision avoidance strategy and guiding the UAV to continue flying toward the target point.
As shown in Fig. 8, in the cross-conflict avoidance simulation, dynamic obstacles invade from the left and right sides of the UAV's flight direction. According to the anti-collision avoidance strategy, when the UAV detects a dynamic obstacle and chooses to pass behind the moving direction of the dynamic obstacle, it can ensure that the UAV and the dynamic obstacle are avoided successfully and the avoidance path is the shortest. The reward function is recalculated after finishing the anti-collision avoidance strategy and guiding the UAV to continue flying toward the target point.
As shown in Fig. 9, in the pursuit conflict avoidance simulation, the flight direction of the dynamic obstacle is the same as the UAV flight direction. Referring to the anti-collision avoidance strategy, when the UAV detects a dynamic obstacle, the UAV chooses to complete the overtake from the right side of the dynamic obstacle's direction of motion, which ensures that the UAV can successfully overtake the dynamic obstacle with the shortest overtake avoidance path; The reward function is recalculated after finishing anti-collision avoidance strategy and guiding the UAV to continue flying towards the target point.
To verify the performance of the APPA-3D, this paper randomly generates 3, 6, and 10 different moving and static obstacles in the same simulation environment and conducts three sets of randomized experiments each. The 3D view of APPA-3D is exhibited in this paper, as shown in Fig. 10.
Graph: Figure 103D view of APPA-3D flight path.
The 3D view of the paths planned by APPA-3D shows that the flight paths of the UAVs are feasible in nine different scenarios. The distance between the UAV and the obstacle is well-maintained in complex terrain. This further demonstrates that APPA-3D can help the UAV to plan a path that is both short and safe at the same time.
This paper calculates four parameters: UAV path planning time, planned path length, number of planned path points, and planned path ground projection length in 9 scenarios, the average values of the four parameters are shown in Table 1.
Table 1 Simulation results in different scenarios.
Number of random obstacles Path planning time Planning path length Number of planning path points Planning path ground projection length 3 32.30 s 14.31083 km 94 11.51960 km 6 45.05 s 14.28151 km 118 13.87523 km 10 51.12 s 13.74009 km 85 10.29533 km
To verify the enhancement effect of the adaptive reward function and the new action selection strategy proposed in this paper, two sets of ablation experiments were conducted firstly before conducting the comparison experiments.
The first set of ablation experiments is to verify the enhancement effect of the new adaptive reward function proposed in this paper, and the experimental results are shown in Fig. 11:
Graph: Figure11UAV path planning results under sparse reward function and adaptive reward function.
The yellow path in Fig. 11 represents the UAV flight path under the sparse reward function, and the blue path represents the flight path under the adaptive reward function proposed in this paper. Figure 11 clearly shows that the performance of UAV path planning based on sparse reward function is poor in complex 3D environments. This is because under sparse reward function, it can only obtain positive and negative reward when reaching the target point or colliding with obstacles, and other actions will not get any positive or negative feedback. So the UAV is random flight blindly, unable to find the correct flight direction in this way. Compared with the sparse reward function, the adaptive reward function we proposed combines the good performance of APF to make the reward accumulation process smoother and can also reflect the relationship between the current state and the target state of the UAV.
To verify the improvement effect of the new action selection strategy proposed in this paper, the second set of ablation experiments was set up. The experiments were analyzed using the
Tables 2 and 3 show the results of three exploration strategies. To prevent the impact of single data on the experiment, the data in Tables 2 and 3 is the average value obtained after 5 experiments.
Table 2 Number of planning path points for different strategies.
Softmax Action selection strategy Number of planning path points 112 96 85
Table 3 Planning path time for different strategies.
Softmax Action selection strategy Planning path time 68.33 s 60.15 s 53.37 s
The experimental results show that after a period of exploration, three different exploration strategies are able to guide the UAV to the target point. Compared with the other two exploration strategies, the action selection probability we proposed is more advantageous in terms of path planning time and number of path planning points.
To evaluate whether the APPA-3D proposed in this paper has significant advantages over other classical or RL based algorithms, two sets of experiments were utilized to test the ability of the six methods to solve path planning problems under the same conditions. According to the characteristics of algorithms, they can be divided into two categories: classic algorithms (APF, RRT, and A*) and QL-based algorithms (DFQL, IQL, and MEAEO-RL). It should be noted that, to prevent the impact of single data on the experiment, the data in Table 4, 5, 6 and 7 is the average value obtained after 5 experiments.
Table 4 Path length of different algorithms.
A* RRT APF APPA-3D Planning path length 15.3205 km 16.5351 km 14.9667 km 13.9533 km
Table 5 Path planning time of different algorithms.
A* RRT APF APPA-3D Planning path time 49.12 s 59.67 s 51.34 s 52.45 s
Table 6 Path lengths of four algorithms.
DFQL IQL MEAEO-RL APPA-3D Planning path length 15.0154 km 14.8301 km 14.6506 km 13.9533 km
Table 7 Path planning time of four algorithms.
DFQL IQL MEAEO-RL APPA-3D Planning path time 54.46 s 57.34 s 54.76 s 52.12 s
The experimental results of the first group are displayed in Fig. 12 and Tables 4 and 5:
Graph: Figure12Path planning comparison of A*, RRT, APF and APPA-3D.
It can be seen from Fig. 12 that the three classic algorithms perform better than APPA-3D algorithm in the front part of the path. However, the performance of classic algorithms is poor in the latter part of the path, which is caused by their algorithm characteristics. Because sampling-based and search-based characteristics of algorithms respectively, it is hard to generate smooth and optimal paths with RRT and A*. Although the path planning time of A* is short, the UAV collided with obstacles unfortunately. The reason for the poor effect of APF is that the obstacle surrounds the destination, and the repulsive force field of the obstacle directly acts on the agent, making it unable to approach the obstacle. The agent can only move in the direction where the gravitational force is greater than the repulsive force.
The results of the second group of experiments are displayed in Fig. 13. It is worth mentioning that DFQL, IQL, MEAEO-R, and APPA-3D are all optimized based on traditional RL algorithms. The simulation results are shown in Fig. 13, Table 6 and Table 7.
Graph: Figure13Path planning comparison of DFQL, IQL, MEAEO-RL and APPA-3D.
Experimental results clearly show that APPA-3D can reach the destination with the shortest distance and time. In the initial phase of path planning, APPA-3D is not very different from other algorithms, and all four algorithms can help the UAV plan a relatively high-quality flight path quickly. While in the middle and later stages of path planning, the differences between APPA-3D and the other three compared algorithms can be seen clearly, especially when facing multi-dynamic obstacles. Because RL algorithm assigns a probability to each possible action and selects the action based on these probabilities, path planning algorithms based on RL often fall into the dilemma of exploration–exploitation when facing complex environments.
To solve this problem, the APPA-3D algorithm proposes a new action selection strategy. This strategy solves the balance problem between exploration and exploitation by introducing the concept of action selection probability and making action preference selection accordingly.
The Fig. 14 presents the loss function used to observe the convergence behavior over iterations of all algorithms. It can be seen that after about 130 iterations, the loss function begins to stabilize. The rapid convergence of value loss also shows that the APPA-3D is more accurate, which is a good performance and means that the agents won't fall into a local optimum. The algorithms compared requires more iteration to complete convergence. This is because they use
Graph: Figure14The loss function of DFQL, IQL, MEAEO-RL and APPA-3D.
In conclusion, APPA-3D is far better than the compared algorithms in the 3D UAV path planning optimization problem. This is because APPA-3D dynamically adjusts the action selection strategy by combining the size of the value function in different states, thus solving the problem of exploration-utilization of RL and improving the efficiency of path search.
The path planning problem in unknown environments is the focus of UAV task planning research and the key to achieving autonomous flight. Therefore, UAVs need to have the ability to autonomous path planning and avoid potential obstacles. In this paper, an autonomous collision-free path planning algorithm for unknown complex 3D environments is proposed. Firstly, based on the environment sensing capability, a UAV collision safety envelope is designed, and the anti-collision control strategy is investigated, which can effectively deal with the collision problem triggered by dynamic obstacles in the flight environment. Secondly, this paper optimizes the traditional RL algorithm. On the one hand, the reward function for RL is optimized by transforming the relationship between the current state of the UAV and the task into a suitable dynamic reward function. The presence of a dynamic reward function allows the UAV to fly toward the target point without getting too close to the obstacles. On the other hand, an RL action exploration strategy based on action selection probability is proposed. The strategy dynamically adjusts the action selection strategy by combining the size of the value function in different states, thus solving the RL exploration-utilization problem and improving the efficiency of path search. To verify the effectiveness of the designed APPA-3D algorithm in the dynamic collision avoidance model, three typical collision experiments were set up, including flight path opposing collision, pursuit collision, and cross collision. The experimental results verify that the APPA-3D can effectively avoid safety threats that may be caused by dynamic obstacles in complex environments according to the designed anti-collision control strategy. Meanwhile, the results of the algorithm testing experiments in nine different scenarios verified that the algorithm still performs well in the face of random and complex flight environments.
APPA-3D demonstrates better performance in path planning performance comparison tests with other classical and novel optimized RL algorithms. The advantages in path planning length and convergence curves again show that APPA-3D can effectively help UAVs solve the path planning problem.
J.W. designed the study and wrote the main manuscript text. Z.Z. experimented and analyzed the results and J.Q. collected the data and analyzed the results. X.C. prepared figures and tables, and all authors reviewed the manuscript.
The datasets used and analysed during the current study will be available from the corresponding author on reasonable request.
The authors declare no competing interests.
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
By Jintao Wang; Zuyi Zhao; Jiayi Qu and Xingguo Chen
Reported by Author; Author; Author; Author