Aggressive and robust low-level control and trajectory tracking for quadrotors with deep reinforcement learning

Executing accurate trajectory tracking tasks using a high-performance low-level controller is crucial for quadrotors to be applied in various scenarios, especially those involving uncertain disturbances. However, due to the uncertainties in disturbed environments, developing effective low-level cont...

Ausführliche Beschreibung

Gespeichert in:
Bibliographische Detailangaben
Veröffentlicht in:Neural computing & applications Jg. 37; H. 3; S. 1223 - 1240
Hauptverfasser: Chen, Shiyu, Li, Yanjie, Lou, Yunjiang, Lin, Ke
Format: Journal Article
Sprache:Englisch
Veröffentlicht: London Springer London 01.01.2025
Springer Nature B.V
Schlagworte:
Accuracy
Artificial Intelligence
Computational Biology/Bioinformatics
Computational Science and Engineering
Computer Science
Control algorithms
Control methods
Controllers
Data Mining and Knowledge Discovery
Deep learning
Disturbances
Hovering
Image Processing and Computer Vision
Initial conditions
Methods
Original Article
Probability and Statistics in Computer Science
Proportional integral derivative
Robotics
Robust control
Rotary wing aircraft
Simulation
Tracking
Unmanned aerial vehicles
Velocity
Vibration control
Robustness
Trajectory tracking
Low-level control
Reinforcement learning
Quadrotor
ISSN:0941-0643, 1433-3058
Online-Zugang:Volltext
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
Beschreibung
Zusammenfassung:Executing accurate trajectory tracking tasks using a high-performance low-level controller is crucial for quadrotors to be applied in various scenarios, especially those involving uncertain disturbances. However, due to the uncertainties in disturbed environments, developing effective low-level controllers with traditional model-based control schemes is challenging. This paper presents an aggressive and robust reinforcement learning (RL)-based low-level control policy for quadrotors. The policy maps the observed quadrotor state directly to motor thrust commands, without requiring the quadrotor dynamics. Additionally, a trajectory generation pipeline is developed to improve the accuracy of trajectory tracking tasks based on differential flatness. With the learned low-level control policy, extensive simulations and real-world experiments are implemented to validate the performance of the policy. The results indicate that our RL-based low-level control policy outperforms traditional proportional–integral–derivative (PID) control methods and related learning-based policies in terms of accuracy and robustness, particularly in environments with uncertain disturbances. Furthermore, the proposed RL-based control policy exhibits an aggressive response in trajectory tracking, even when the speed of the desired trajectory is increased to 6 m/s. Moreover, the learned policy demonstrates strong vibration suppression capabilities and enables the quadrotor to recover to a hovering state from random initial conditions with shorter response time.
Bibliographie:ObjectType-Article-1
SourceType-Scholarly Journals-1
ObjectType-Feature-2
content type line 14
ISSN:0941-0643
1433-3058
DOI:10.1007/s00521-024-10675-4